From 3404379a62711d102f448052a1cbb98d5ea95fbc Mon Sep 17 00:00:00 2001
From: Wenyao Gao <wgao11@u.rochester.edu>
Date: Wed, 22 Jul 2026 13:43:11 -0700
Subject: [PATCH 01/13] feat(overview): add executive overview with
 precision-aware rankings and result-level evidence

---
 .prettierignore                               |   3 +
 packages/app/cypress/component/header.cy.tsx  | 101 ++
 packages/app/cypress/e2e/overview.cy.ts       | 151 +++
 .../cypress/fixtures/api/overview-rows.json   |  26 +
 packages/app/src/app/overview/page.tsx        |  40 +
 packages/app/src/app/sitemap.ts               |   5 +
 packages/app/src/app/zh/overview/page.tsx     |  41 +
 packages/app/src/components/header/header.tsx |  28 +-
 .../overview/overview-detail-link.tsx         |  86 ++
 .../src/components/overview/overview-page.tsx |  59 ++
 .../overview/overview-scorecard.tsx           | 736 +++++++++++++
 .../app/src/components/ui/mode-toggle.tsx     |   2 +-
 packages/app/src/hooks/useTogglableSet.ts     |  29 +-
 packages/app/src/lib/benchmark-data.server.ts |  21 +
 packages/app/src/lib/compare-ssr.ts           |  27 +-
 packages/app/src/lib/exclusion.ts             |   2 +-
 packages/app/src/lib/i18n.test.ts             |   3 +
 packages/app/src/lib/i18n.ts                  |   1 +
 .../src/lib/overview-config-identity.test.ts  |  97 ++
 .../app/src/lib/overview-config-identity.ts   |  33 +
 packages/app/src/lib/overview-data.server.ts  |  29 +
 packages/app/src/lib/overview-data.test.ts    | 740 +++++++++++++
 packages/app/src/lib/overview-data.ts         | 969 ++++++++++++++++++
 packages/app/src/lib/overview-links.test.ts   | 100 ++
 packages/app/src/lib/overview-links.ts        |  89 ++
 packages/app/src/lib/tco-feed.ts              |  36 +
 packages/app/src/lib/toggle-set.ts            |  32 +
 27 files changed, 3427 insertions(+), 59 deletions(-)
 create mode 100644 .prettierignore
 create mode 100644 packages/app/cypress/e2e/overview.cy.ts
 create mode 100644 packages/app/cypress/fixtures/api/overview-rows.json
 create mode 100644 packages/app/src/app/overview/page.tsx
 create mode 100644 packages/app/src/app/zh/overview/page.tsx
 create mode 100644 packages/app/src/components/overview/overview-detail-link.tsx
 create mode 100644 packages/app/src/components/overview/overview-page.tsx
 create mode 100644 packages/app/src/components/overview/overview-scorecard.tsx
 create mode 100644 packages/app/src/lib/benchmark-data.server.ts
 create mode 100644 packages/app/src/lib/overview-config-identity.test.ts
 create mode 100644 packages/app/src/lib/overview-config-identity.ts
 create mode 100644 packages/app/src/lib/overview-data.server.ts
 create mode 100644 packages/app/src/lib/overview-data.test.ts
 create mode 100644 packages/app/src/lib/overview-data.ts
 create mode 100644 packages/app/src/lib/overview-links.test.ts
 create mode 100644 packages/app/src/lib/overview-links.ts
 create mode 100644 packages/app/src/lib/toggle-set.ts

diff --git a/.prettierignore b/.prettierignore
new file mode 100644
index 00000000..3c258190
--- /dev/null
+++ b/.prettierignore
@@ -0,0 +1,3 @@
+# Deliberately compact test fixture (one benchmark row per line, <=60 lines);
+# oxfmt/prettier would pretty-print it to ~530 lines. See overview-data.test.ts drift guard.
+packages/app/cypress/fixtures/api/overview-rows.json
diff --git a/packages/app/cypress/component/header.cy.tsx b/packages/app/cypress/component/header.cy.tsx
index 747396ea..308b31d4 100644
--- a/packages/app/cypress/component/header.cy.tsx
+++ b/packages/app/cypress/component/header.cy.tsx
@@ -5,10 +5,24 @@ import { PathnameContext } from 'next/dist/shared/lib/hooks-client-context.share
 import { Header } from '@/components/header/header';
 import { ThemeProvider } from '@/components/ui/theme-provider';
 
+// Component specs mount outside the Next app shell, so the global stylesheet is
+// never injected. The layout assertions below need real Tailwind output, and
+// next-style-loader inserts before this anchor node, so both must exist before
+// the stylesheet module evaluates.
+const cssAnchor = document.createElement('noscript');
+cssAnchor.id = '__next_css__DO_NOT_USE__';
+document.head.append(cssAnchor);
+require('@/app/globals.css');
+
 const queryClient = new QueryClient({
   defaultOptions: { queries: { retry: false } },
 });
 
+/** Minimum touch target, per WCAG 2.5.8 and the `size-11` / `min-h-11` utilities. */
+const MIN_TOUCH_PX = 44;
+/** Tolerance for sub-pixel layout rounding. */
+const EPSILON = 0.5;
+
 function createMockRouter() {
   return {
     push: cy.stub(),
@@ -20,6 +34,10 @@ function createMockRouter() {
   };
 }
 
+function rectOf(selector: string) {
+  return cy.get(selector).then(($el) => $el[0].getBoundingClientRect());
+}
+
 describe('Header', () => {
   beforeEach(() => {
     const mockRouter = createMockRouter();
@@ -84,4 +102,87 @@ describe('Header', () => {
       cy.contains('a', 'Supporters').should('be.visible').and('have.attr', 'href', '/quotes');
     });
   });
+
+  describe('at 320x700', () => {
+    beforeEach(() => {
+      cy.viewport(320, 700);
+    });
+
+    it('hides the GitHub star control', () => {
+      cy.get('[data-testid="header-star-button"]').should('not.be.visible');
+    });
+
+    it('keeps the remaining controls inside the header bounds', () => {
+      cy.get('[data-testid="header"]').then(($header) => {
+        const bounds = $header[0].getBoundingClientRect();
+        const selectors = [
+          '[data-testid="header-brand"]',
+          '[data-testid="language-toggle"]',
+          '[data-testid="theme-toggle"]',
+          '[data-testid="mobile-menu-toggle"]',
+        ];
+        selectors.forEach((selector) => {
+          rectOf(selector).then((rect) => {
+            expect(rect.left, `${selector} left edge`).to.be.at.least(bounds.left - EPSILON);
+            expect(rect.right, `${selector} right edge`).to.be.at.most(bounds.right + EPSILON);
+          });
+        });
+      });
+    });
+
+    it('gives the brand and language controls a 44px touch height', () => {
+      ['[data-testid="header-brand"]', '[data-testid="language-toggle"]'].forEach((selector) => {
+        rectOf(selector).then((rect) => {
+          expect(rect.height, `${selector} height`).to.be.at.least(MIN_TOUCH_PX - EPSILON);
+        });
+      });
+    });
+
+    it('gives the icon controls a 44px touch target in both dimensions', () => {
+      ['[data-testid="theme-toggle"]', '[data-testid="mobile-menu-toggle"]'].forEach((selector) => {
+        rectOf(selector).then((rect) => {
+          expect(rect.width, `${selector} width`).to.be.at.least(MIN_TOUCH_PX - EPSILON);
+          expect(rect.height, `${selector} height`).to.be.at.least(MIN_TOUCH_PX - EPSILON);
+        });
+      });
+    });
+
+    it('does not overflow horizontally', () => {
+      cy.get('[data-testid="header"]').then(($header) => {
+        const header = $header[0];
+        expect(header.scrollWidth, 'header scrollWidth').to.be.at.most(header.clientWidth);
+      });
+    });
+
+    it('still opens the menu and exposes its links', () => {
+      cy.get('[data-testid="mobile-menu-toggle"]').click();
+      cy.get('[data-testid="mobile-menu"]').should('be.visible');
+      cy.get('[data-testid="mobile-menu"]').within(() => {
+        ['Home', 'Dashboard', 'Comparisons', 'Supporters', 'Datasets', 'Articles', 'About'].forEach(
+          (label) => {
+            cy.contains('a', label).should('be.visible');
+          },
+        );
+      });
+      cy.get('[data-testid="mobile-menu"] a').each(($link) => {
+        const rect = $link[0].getBoundingClientRect();
+        expect(rect.height, `${$link.text()} link height`).to.be.at.least(MIN_TOUCH_PX - EPSILON);
+      });
+    });
+
+    it('exposes the minecraft audio toggles in the mobile menu without overflowing', () => {
+      cy.get('[data-testid="theme-toggle"]').click();
+      cy.get('[data-testid="theme-toggle"]').click();
+      cy.get('html').should('have.class', 'minecraft');
+      cy.get('[data-testid="mobile-menu-toggle"]').click();
+      cy.get('[data-testid="mobile-menu"]').within(() => {
+        cy.get('button[aria-label="Mute music"]').should('be.visible');
+        cy.get('button[aria-label="Mute click sounds"]').should('be.visible');
+      });
+      cy.get('[data-testid="header"]').then(($header) => {
+        const header = $header[0];
+        expect(header.scrollWidth, 'header scrollWidth').to.be.at.most(header.clientWidth);
+      });
+    });
+  });
 });
diff --git a/packages/app/cypress/e2e/overview.cy.ts b/packages/app/cypress/e2e/overview.cy.ts
new file mode 100644
index 00000000..bdd92f69
--- /dev/null
+++ b/packages/app/cypress/e2e/overview.cy.ts
@@ -0,0 +1,151 @@
+// Focused smoke coverage for /overview. In fixtures mode the page is served by
+// the synthetic cypress/fixtures/api/overview-rows.json through the real data
+// builder, so every expected value below was derived by running the assembler
+// over that fixture (the drift guard in overview-data.test.ts locks the same
+// values) — never eyeballed. The fixture exercises every 8.7 state: coverage-driven
+// primary precision, ranked vs coverage secondary, all not-ranked reasons, and —
+// uniquely — a cross-day evidence range the live dataset does not produce.
+
+const MODEL_LABELS = [
+  'DeepSeek V4 Pro 1.6T',
+  'Kimi K2.5/2.6/2.7-Code 1T',
+  'MiniMax M3 428B',
+  'GLM5.2',
+  'Qwen3.5 397B',
+];
+
+/** The page must never scroll sideways: the whole comparison has to fit. */
+function expectNoHorizontalOverflow() {
+  cy.document().then((doc) => {
+    expect(doc.documentElement.scrollWidth).to.be.lte(doc.documentElement.clientWidth);
+  });
+}
+
+/**
+ * A narrow document is not enough: a table parked in an `overflow-x` region
+ * keeps the document narrow while still scrolling sideways, so nothing inside
+ * the surface may be wider than its own box either. `sr-only` clips (1px) and
+ * non-replaced inline boxes are exempt — CSSOM reports a spurious `scrollWidth`
+ * for inline boxes in Firefox, and `overflow` cannot apply to them anyway.
+ */
+function expectNoHorizontalScroller(testId: string) {
+  cy.get(`[data-testid="${testId}"]`).then(([surface]) => {
+    const scrollers = [surface, ...surface.querySelectorAll('*')]
+      .filter(
+        (el) =>
+          !el.classList.contains('sr-only') &&
+          getComputedStyle(el).display !== 'inline' &&
+          el.scrollWidth > el.clientWidth + 1,
+      )
+      .map((el) => `${el.tagName} ${el.scrollWidth}>${el.clientWidth}`);
+    expect(scrollers, `horizontally scrollable inside ${testId}`).to.deep.equal([]);
+  });
+}
+
+function desktopModel(model: string) {
+  return cy.get(`[data-testid="overview-desktop-model"][data-model="${model}"]`);
+}
+
+describe('Overview page', () => {
+  it('summarizes every active model without interactive widgets', () => {
+    cy.viewport(1280, 800);
+    cy.visit('/overview');
+
+    cy.contains('h1', 'AI Inference Overview').should('exist');
+    cy.get('[data-testid="overview-desktop-matrix"]').should('be.visible');
+    cy.get('[data-testid="overview-desktop-matrix"] h2').should('have.length', MODEL_LABELS.length);
+    for (const label of MODEL_LABELS) {
+      cy.get('[data-testid="overview-desktop-matrix"]').should('contain.text', label);
+    }
+    // Database-wide freshness line + the methodology footnote's once-per-page note.
+    cy.contains('Database snapshot through Jul 18').should('exist');
+    cy.contains(
+      'The default precision maximizes comparable hardware coverage. Not ranked does not mean slower.',
+    ).should('exist');
+    // A one-glance summary, not an interactive widget.
+    cy.get('[data-testid="overview-desktop-matrix"]').within(() => {
+      cy.get('details, summary, button').should('not.exist');
+    });
+  });
+
+  it('ranks the primary precision, dates each read, and accounts for every hardware', () => {
+    cy.viewport(1280, 800);
+    cy.visit('/overview');
+
+    desktopModel('DeepSeek-V4-Pro').within(() => {
+      // Primary precision, its leader, and the runner-up's own signed delta (ASCII minus).
+      cy.contains('FP4 · PRIMARY @50').should('exist');
+      cy.contains('B300').should('exist');
+      cy.contains('1,122').should('exist');
+      cy.contains('Leader').should('exist');
+      cy.contains('-20%').should('exist');
+      // FP8 measured one platform → one compact coverage line, not a second table.
+      cy.contains('FP8 coverage: GB200 NVL72 measured; insufficient comparable results.').should(
+        'exist',
+      );
+      // Leader @50 is bracketed by two run days (en-dash range); same-day reads and
+      // the @100 read (its own leader) collapse to a single date.
+      cy.contains('Jun 24–Jul 4').should('exist');
+      cy.contains('Jul 18').should('exist');
+      cy.contains('381').should('exist');
+      cy.contains('Only exact result').should('exist');
+      // Every remaining hardware carries one reason, both clamp directions included.
+      cy.contains('cannot reach @50').should('exist');
+      cy.contains('no exact @50 result').should('exist');
+      cy.contains('no 8K/1K data').should('exist');
+      cy.contains('standard decode only').should('exist');
+      // Each ranked value links into its own pre-filtered dashboard view.
+      cy.get('a[href*="i_gpus="]')
+        .first()
+        .should('have.attr', 'href')
+        .and('include', 'g_model=DeepSeek-V4-Pro')
+        .and('include', 'i_gpus=b300_sglang_mtp')
+        .and('include', 'i_spec=mtp');
+    });
+  });
+
+  it('opens a ranked secondary only when the other precision adds comparable hardware', () => {
+    cy.viewport(1280, 800);
+    cy.visit('/overview');
+
+    // Qwen: FP8 ranks a comparable pair AND adds MI355X, so it renders subordinate rows.
+    desktopModel('Qwen-3.5-397B-A17B').within(() => {
+      cy.contains('FP4 · PRIMARY @50').should('exist');
+      cy.contains('FP8 @50').should('exist');
+      cy.contains('MI355X').should('exist');
+      cy.contains('760').should('exist');
+      cy.contains('-16%').should('exist');
+    });
+    // MiniMax: wider exact-@50 FP8 coverage flips the primary precision to FP8;
+    // FP4 falls back to a single-hardware coverage line.
+    desktopModel('MiniMax-M3').within(() => {
+      cy.contains('FP8 · PRIMARY @50').should('exist');
+      cy.contains('FP4 coverage: H200 measured; insufficient comparable results.').should('exist');
+    });
+  });
+
+  it('stacks on a 390px phone with no sideways scroll', () => {
+    cy.viewport(390, 844);
+    cy.visit('/overview');
+
+    cy.get('[data-testid="overview-mobile-list"]').should('be.visible');
+    cy.get('[data-testid="overview-desktop-matrix"]').should('not.be.visible');
+    cy.get('[data-testid="overview-mobile-list"]').within(() => {
+      cy.get('details, summary, button').should('not.exist');
+    });
+    expectNoHorizontalOverflow();
+    expectNoHorizontalScroller('overview-mobile-list');
+  });
+
+  it('renders the Chinese sibling with the same hierarchy and the per-GPU unit', () => {
+    cy.viewport(1280, 800);
+    cy.visit('/zh/overview');
+
+    cy.contains('h1', 'AI 推理总览').should('exist');
+    cy.contains('输出 tok/s/GPU').should('exist');
+    desktopModel('DeepSeek-V4-Pro').within(() => {
+      cy.contains('FP4 · 主排名 @50').should('exist');
+    });
+    cy.contains('默认精度优先覆盖更多可比较硬件。未参与排名不代表性能更低。').should('exist');
+  });
+});
diff --git a/packages/app/cypress/fixtures/api/overview-rows.json b/packages/app/cypress/fixtures/api/overview-rows.json
new file mode 100644
index 00000000..d0e53d8a
--- /dev/null
+++ b/packages/app/cypress/fixtures/api/overview-rows.json
@@ -0,0 +1,26 @@
+{
+  "DeepSeek-V4-Pro": [
+    {"id":1,"hardware":"b300","framework":"sglang","model":"dsv4","precision":"fp4","spec_method":"mtp","disagg":false,"is_multinode":false,"prefill_tp":8,"prefill_ep":1,"prefill_dp_attention":false,"prefill_num_workers":1,"decode_tp":8,"decode_ep":1,"decode_dp_attention":false,"decode_num_workers":1,"num_prefill_gpu":8,"num_decode_gpu":8,"benchmark_type":"single_turn","isl":8192,"osl":1024,"conc":8,"offload_mode":"off","image":null,"metrics":{"median_intvty":30,"output_tput_per_gpu":1300},"date":"2026-06-24","run_url":null},
+    {"id":2,"hardware":"b300","framework":"sglang","model":"dsv4","precision":"fp4","spec_method":"mtp","disagg":false,"is_multinode":false,"prefill_tp":8,"prefill_ep":1,"prefill_dp_attention":false,"prefill_num_workers":1,"decode_tp":8,"decode_ep":1,"decode_dp_attention":false,"decode_num_workers":1,"num_prefill_gpu":8,"num_decode_gpu":8,"benchmark_type":"single_turn","isl":8192,"osl":1024,"conc":16,"offload_mode":"off","image":null,"metrics":{"median_intvty":70,"output_tput_per_gpu":1000},"date":"2026-07-04","run_url":null},
+    {"id":3,"hardware":"b200","framework":"sglang","model":"dsv4","precision":"fp4","spec_method":"mtp","disagg":false,"is_multinode":false,"prefill_tp":8,"prefill_ep":1,"prefill_dp_attention":false,"prefill_num_workers":1,"decode_tp":8,"decode_ep":1,"decode_dp_attention":false,"decode_num_workers":1,"num_prefill_gpu":8,"num_decode_gpu":8,"benchmark_type":"single_turn","isl":8192,"osl":1024,"conc":8,"offload_mode":"off","image":null,"metrics":{"median_intvty":40,"output_tput_per_gpu":1050},"date":"2026-07-18","run_url":null},
+    {"id":4,"hardware":"b200","framework":"sglang","model":"dsv4","precision":"fp4","spec_method":"mtp","disagg":false,"is_multinode":false,"prefill_tp":8,"prefill_ep":1,"prefill_dp_attention":false,"prefill_num_workers":1,"decode_tp":8,"decode_ep":1,"decode_dp_attention":false,"decode_num_workers":1,"num_prefill_gpu":8,"num_decode_gpu":8,"benchmark_type":"single_turn","isl":8192,"osl":1024,"conc":24,"offload_mode":"off","image":null,"metrics":{"median_intvty":110,"output_tput_per_gpu":300},"date":"2026-07-18","run_url":null},
+    {"id":5,"hardware":"mi355x","framework":"sglang","model":"dsv4","precision":"fp4","spec_method":"mtp","disagg":false,"is_multinode":false,"prefill_tp":8,"prefill_ep":1,"prefill_dp_attention":false,"prefill_num_workers":1,"decode_tp":8,"decode_ep":1,"decode_dp_attention":false,"decode_num_workers":1,"num_prefill_gpu":8,"num_decode_gpu":8,"benchmark_type":"single_turn","isl":8192,"osl":1024,"conc":16,"offload_mode":"off","image":null,"metrics":{"median_intvty":60,"output_tput_per_gpu":760},"date":"2026-07-18","run_url":null},
+    {"id":6,"hardware":"gb300","framework":"sglang","model":"dsv4","precision":"fp4","spec_method":"mtp","disagg":false,"is_multinode":false,"prefill_tp":8,"prefill_ep":1,"prefill_dp_attention":false,"prefill_num_workers":1,"decode_tp":8,"decode_ep":1,"decode_dp_attention":false,"decode_num_workers":1,"num_prefill_gpu":8,"num_decode_gpu":8,"benchmark_type":"single_turn","isl":8192,"osl":1024,"conc":8,"offload_mode":"off","image":null,"metrics":{"median_intvty":45,"output_tput_per_gpu":1050},"date":"2026-07-18","run_url":null},
+    {"id":7,"hardware":"h100","framework":"sglang","model":"dsv4","precision":"fp4","spec_method":"none","disagg":false,"is_multinode":false,"prefill_tp":8,"prefill_ep":1,"prefill_dp_attention":false,"prefill_num_workers":1,"decode_tp":8,"decode_ep":1,"decode_dp_attention":false,"decode_num_workers":1,"num_prefill_gpu":8,"num_decode_gpu":8,"benchmark_type":"single_turn","isl":8192,"osl":1024,"conc":8,"offload_mode":"off","image":null,"metrics":{"median_intvty":30,"output_tput_per_gpu":800},"date":"2026-07-18","run_url":null},
+    {"id":8,"hardware":"gb200","framework":"sglang","model":"dsv4","precision":"fp8","spec_method":"mtp","disagg":false,"is_multinode":false,"prefill_tp":8,"prefill_ep":1,"prefill_dp_attention":false,"prefill_num_workers":1,"decode_tp":8,"decode_ep":1,"decode_dp_attention":false,"decode_num_workers":1,"num_prefill_gpu":8,"num_decode_gpu":8,"benchmark_type":"single_turn","isl":8192,"osl":1024,"conc":8,"offload_mode":"off","image":null,"metrics":{"median_intvty":50,"output_tput_per_gpu":600},"date":"2026-07-18","run_url":null}
+  ],
+  "Qwen-3.5-397B-A17B": [
+    {"id":9,"hardware":"b300","framework":"sglang","model":"qwen3.5","precision":"fp4","spec_method":"mtp","disagg":false,"is_multinode":false,"prefill_tp":8,"prefill_ep":1,"prefill_dp_attention":false,"prefill_num_workers":1,"decode_tp":8,"decode_ep":1,"decode_dp_attention":false,"decode_num_workers":1,"num_prefill_gpu":8,"num_decode_gpu":8,"benchmark_type":"single_turn","isl":8192,"osl":1024,"conc":8,"offload_mode":"off","image":null,"metrics":{"median_intvty":30,"output_tput_per_gpu":1400},"date":"2026-07-18","run_url":null},
+    {"id":10,"hardware":"b300","framework":"sglang","model":"qwen3.5","precision":"fp4","spec_method":"mtp","disagg":false,"is_multinode":false,"prefill_tp":8,"prefill_ep":1,"prefill_dp_attention":false,"prefill_num_workers":1,"decode_tp":8,"decode_ep":1,"decode_dp_attention":false,"decode_num_workers":1,"num_prefill_gpu":8,"num_decode_gpu":8,"benchmark_type":"single_turn","isl":8192,"osl":1024,"conc":16,"offload_mode":"off","image":null,"metrics":{"median_intvty":70,"output_tput_per_gpu":980},"date":"2026-07-18","run_url":null},
+    {"id":11,"hardware":"b200","framework":"sglang","model":"qwen3.5","precision":"fp4","spec_method":"mtp","disagg":false,"is_multinode":false,"prefill_tp":8,"prefill_ep":1,"prefill_dp_attention":false,"prefill_num_workers":1,"decode_tp":8,"decode_ep":1,"decode_dp_attention":false,"decode_num_workers":1,"num_prefill_gpu":8,"num_decode_gpu":8,"benchmark_type":"single_turn","isl":8192,"osl":1024,"conc":8,"offload_mode":"off","image":null,"metrics":{"median_intvty":40,"output_tput_per_gpu":1150},"date":"2026-07-18","run_url":null},
+    {"id":12,"hardware":"b200","framework":"sglang","model":"qwen3.5","precision":"fp4","spec_method":"mtp","disagg":false,"is_multinode":false,"prefill_tp":8,"prefill_ep":1,"prefill_dp_attention":false,"prefill_num_workers":1,"decode_tp":8,"decode_ep":1,"decode_dp_attention":false,"decode_num_workers":1,"num_prefill_gpu":8,"num_decode_gpu":8,"benchmark_type":"single_turn","isl":8192,"osl":1024,"conc":24,"offload_mode":"off","image":null,"metrics":{"median_intvty":80,"output_tput_per_gpu":720},"date":"2026-07-18","run_url":null},
+    {"id":13,"hardware":"b200","framework":"sglang","model":"qwen3.5","precision":"fp8","spec_method":"mtp","disagg":false,"is_multinode":false,"prefill_tp":8,"prefill_ep":1,"prefill_dp_attention":false,"prefill_num_workers":1,"decode_tp":8,"decode_ep":1,"decode_dp_attention":false,"decode_num_workers":1,"num_prefill_gpu":8,"num_decode_gpu":8,"benchmark_type":"single_turn","isl":8192,"osl":1024,"conc":12,"offload_mode":"off","image":null,"metrics":{"median_intvty":50,"output_tput_per_gpu":900},"date":"2026-07-18","run_url":null},
+    {"id":14,"hardware":"mi355x","framework":"sglang","model":"qwen3.5","precision":"fp8","spec_method":"mtp","disagg":false,"is_multinode":false,"prefill_tp":8,"prefill_ep":1,"prefill_dp_attention":false,"prefill_num_workers":1,"decode_tp":8,"decode_ep":1,"decode_dp_attention":false,"decode_num_workers":1,"num_prefill_gpu":8,"num_decode_gpu":8,"benchmark_type":"single_turn","isl":8192,"osl":1024,"conc":12,"offload_mode":"off","image":null,"metrics":{"median_intvty":50,"output_tput_per_gpu":760},"date":"2026-07-18","run_url":null}
+  ],
+  "MiniMax-M3": [
+    {"id":15,"hardware":"h200","framework":"sglang","model":"minimaxm3","precision":"fp8","spec_method":"mtp","disagg":false,"is_multinode":false,"prefill_tp":8,"prefill_ep":1,"prefill_dp_attention":false,"prefill_num_workers":1,"decode_tp":8,"decode_ep":1,"decode_dp_attention":false,"decode_num_workers":1,"num_prefill_gpu":8,"num_decode_gpu":8,"benchmark_type":"single_turn","isl":8192,"osl":1024,"conc":8,"offload_mode":"off","image":null,"metrics":{"median_intvty":30,"output_tput_per_gpu":820},"date":"2026-07-18","run_url":null},
+    {"id":16,"hardware":"h200","framework":"sglang","model":"minimaxm3","precision":"fp8","spec_method":"mtp","disagg":false,"is_multinode":false,"prefill_tp":8,"prefill_ep":1,"prefill_dp_attention":false,"prefill_num_workers":1,"decode_tp":8,"decode_ep":1,"decode_dp_attention":false,"decode_num_workers":1,"num_prefill_gpu":8,"num_decode_gpu":8,"benchmark_type":"single_turn","isl":8192,"osl":1024,"conc":16,"offload_mode":"off","image":null,"metrics":{"median_intvty":70,"output_tput_per_gpu":560},"date":"2026-07-18","run_url":null},
+    {"id":17,"hardware":"mi325x","framework":"sglang","model":"minimaxm3","precision":"fp8","spec_method":"mtp","disagg":false,"is_multinode":false,"prefill_tp":8,"prefill_ep":1,"prefill_dp_attention":false,"prefill_num_workers":1,"decode_tp":8,"decode_ep":1,"decode_dp_attention":false,"decode_num_workers":1,"num_prefill_gpu":8,"num_decode_gpu":8,"benchmark_type":"single_turn","isl":8192,"osl":1024,"conc":12,"offload_mode":"off","image":null,"metrics":{"median_intvty":50,"output_tput_per_gpu":650},"date":"2026-07-18","run_url":null},
+    {"id":18,"hardware":"h200","framework":"sglang","model":"minimaxm3","precision":"fp4","spec_method":"mtp","disagg":false,"is_multinode":false,"prefill_tp":8,"prefill_ep":1,"prefill_dp_attention":false,"prefill_num_workers":1,"decode_tp":8,"decode_ep":1,"decode_dp_attention":false,"decode_num_workers":1,"num_prefill_gpu":8,"num_decode_gpu":8,"benchmark_type":"single_turn","isl":8192,"osl":1024,"conc":12,"offload_mode":"off","image":null,"metrics":{"median_intvty":50,"output_tput_per_gpu":600},"date":"2026-07-18","run_url":null}
+  ]
+}
diff --git a/packages/app/src/app/overview/page.tsx b/packages/app/src/app/overview/page.tsx
new file mode 100644
index 00000000..e76d0d35
--- /dev/null
+++ b/packages/app/src/app/overview/page.tsx
@@ -0,0 +1,40 @@
+import type { Metadata } from 'next';
+
+import { SITE_NAME, SITE_URL } from '@semianalysisai/inferencex-constants';
+
+import { OverviewPageContent } from '@/components/overview/overview-page';
+import { enAlternates } from '@/lib/i18n';
+import { getOverviewPageData } from '@/lib/overview-data.server';
+
+export const dynamic = 'force-dynamic';
+
+const DESCRIPTION =
+  'Comparable validated AI inference serving results for every active model and platform at a fixed single-turn 8K input / 1K output workload, ranked at 50 tok/s/user.';
+
+export const metadata: Metadata = {
+  title: 'AI Inference Overview',
+  description: DESCRIPTION,
+  alternates: enAlternates('/overview'),
+  openGraph: {
+    title: `AI Inference Overview | ${SITE_NAME}`,
+    description: DESCRIPTION,
+    url: `${SITE_URL}/overview`,
+    type: 'website',
+  },
+  twitter: {
+    card: 'summary_large_image',
+    title: `AI Inference Overview | ${SITE_NAME}`,
+    description: DESCRIPTION,
+  },
+};
+
+export default async function OverviewPage() {
+  const data = await getOverviewPageData();
+  return (
+    <main className="relative">
+      <div className="container mx-auto px-4 lg:px-8 pb-8">
+        <OverviewPageContent data={data} locale="en" />
+      </div>
+    </main>
+  );
+}
diff --git a/packages/app/src/app/sitemap.ts b/packages/app/src/app/sitemap.ts
index 32bc84d4..2dd3e934 100644
--- a/packages/app/src/app/sitemap.ts
+++ b/packages/app/src/app/sitemap.ts
@@ -58,6 +58,11 @@ export default async function sitemap(): Promise<MetadataRoute.Sitemap> {
 
   return [
     ...localizedPair('/', { lastModified: now, changeFrequency: 'daily', priority: 1 }),
+    ...localizedPair('/overview', {
+      lastModified: now,
+      changeFrequency: 'daily',
+      priority: 0.9,
+    }),
     ...TABS.flatMap((tab) =>
       localizedPair(`/${tab}`, {
         lastModified: now,
diff --git a/packages/app/src/app/zh/overview/page.tsx b/packages/app/src/app/zh/overview/page.tsx
new file mode 100644
index 00000000..91e4fd48
--- /dev/null
+++ b/packages/app/src/app/zh/overview/page.tsx
@@ -0,0 +1,41 @@
+import type { Metadata } from 'next';
+
+import { SITE_NAME, SITE_URL } from '@semianalysisai/inferencex-constants';
+
+import { OverviewPageContent } from '@/components/overview/overview-page';
+import { ZH_OG_LOCALE, zhAlternates } from '@/lib/i18n';
+import { getOverviewPageData } from '@/lib/overview-data.server';
+
+export const dynamic = 'force-dynamic';
+
+const DESCRIPTION =
+  '以固定单轮 8K 输入 / 1K 输出负载，按 50 tok/s/user 档位排名，对比每个活跃模型在各硬件平台上可比的已验证服务结果。';
+
+export const metadata: Metadata = {
+  title: 'AI 推理总览',
+  description: DESCRIPTION,
+  alternates: zhAlternates('/overview'),
+  openGraph: {
+    title: `AI 推理总览 | ${SITE_NAME}`,
+    description: DESCRIPTION,
+    url: `${SITE_URL}/zh/overview`,
+    type: 'website',
+    locale: ZH_OG_LOCALE,
+  },
+  twitter: {
+    card: 'summary_large_image',
+    title: `AI 推理总览 | ${SITE_NAME}`,
+    description: DESCRIPTION,
+  },
+};
+
+export default async function ZhOverviewPage() {
+  const data = await getOverviewPageData();
+  return (
+    <main className="relative">
+      <div className="container mx-auto px-4 lg:px-8 pb-8">
+        <OverviewPageContent data={data} locale="zh" />
+      </div>
+    </main>
+  );
+}
diff --git a/packages/app/src/components/header/header.tsx b/packages/app/src/components/header/header.tsx
index b5fbbf71..09926859 100644
--- a/packages/app/src/components/header/header.tsx
+++ b/packages/app/src/components/header/header.tsx
@@ -83,7 +83,7 @@ function LanguageToggle({ pathname }: { pathname: string }) {
       href={target}
       data-testid="language-toggle"
       hrefLang={isZh ? 'en' : 'zh-CN'}
-      className="px-2 py-1.5 rounded-md text-sm font-medium text-muted-foreground hover:text-foreground hover:bg-muted transition-colors whitespace-nowrap"
+      className="inline-flex items-center min-h-11 px-2 rounded-md text-sm font-medium text-muted-foreground hover:text-foreground hover:bg-muted transition-colors whitespace-nowrap"
       onClick={() => track('header_language_toggled', { to: isZh ? 'en' : 'zh' })}
     >
       {isZh ? 'EN' : '中文'}
@@ -145,7 +145,11 @@ export const Header = ({ starCount }: { starCount?: number | null }) => {
       <div className="container mx-auto px-4 lg:px-8">
         <div className="flex h-14 items-center gap-6">
           {/* Brand */}
-          <Link href={isZh ? '/zh' : '/'} className="flex items-center gap-2 shrink-0">
+          <Link
+            href={isZh ? '/zh' : '/'}
+            data-testid="header-brand"
+            className="flex items-center min-h-11 gap-2 shrink-0"
+          >
             <span className="pride-wordmark text-lg font-bold tracking-tight">InferenceX</span>
             <span className="hidden sm:flex items-center gap-1.5 text-xs text-muted-foreground">
               by
@@ -184,9 +188,18 @@ export const Header = ({ starCount }: { starCount?: number | null }) => {
 
           {/* Right side */}
           <div className="ml-auto flex items-center gap-2">
-            <GitHubStars owner="SemiAnalysisAI" repo="InferenceX" starCount={starCount} />
+            <span className="hidden sm:flex">
+              <GitHubStars owner="SemiAnalysisAI" repo="InferenceX" starCount={starCount} />
+            </span>
             <LanguageToggle pathname={pathname} />
-            <MinecraftToggles />
+            {/* Minecraft-only audio toggles ride alongside GitHub stars below the
+                same `sm` cutoff: in minecraft mode the pixel font widens the bar,
+                and these two extra buttons are what push a 320px header past its
+                bounds. Below `sm` they move into the mobile menu instead (see the
+                MinecraftToggles instance in the hamburger dropdown below). */}
+            <span className="hidden items-center gap-2 sm:flex">
+              <MinecraftToggles />
+            </span>
             <ModeToggle />
 
             {/* Mobile hamburger */}
@@ -195,7 +208,7 @@ export const Header = ({ starCount }: { starCount?: number | null }) => {
                 type="button"
                 data-testid="mobile-menu-toggle"
                 onClick={toggleMenu}
-                className="flex items-center justify-center size-9 rounded-md transition-colors hover:bg-muted cursor-pointer"
+                className="flex items-center justify-center size-11 rounded-md transition-colors hover:bg-muted cursor-pointer"
                 aria-expanded={mobileMenuOpen}
                 aria-label="Navigation menu"
               >
@@ -224,7 +237,7 @@ export const Header = ({ starCount }: { starCount?: number | null }) => {
                       key={href}
                       href={displayHref}
                       className={cn(
-                        'px-3 py-2 rounded-md text-sm font-medium transition-colors',
+                        'flex items-center min-h-11 px-3 rounded-md text-sm font-medium transition-colors',
                         isActive(pathname, href)
                           ? 'text-brand bg-brand/10'
                           : 'text-muted-foreground hover:text-foreground hover:bg-muted',
@@ -237,6 +250,9 @@ export const Header = ({ starCount }: { starCount?: number | null }) => {
                       {label}
                     </Link>
                   ))}
+                  <span className="flex items-center gap-2 px-3 sm:hidden">
+                    <MinecraftToggles />
+                  </span>
                 </div>
               )}
             </div>
diff --git a/packages/app/src/components/overview/overview-detail-link.tsx b/packages/app/src/components/overview/overview-detail-link.tsx
new file mode 100644
index 00000000..f7b23298
--- /dev/null
+++ b/packages/app/src/components/overview/overview-detail-link.tsx
@@ -0,0 +1,86 @@
+'use client';
+
+import { ArrowRight } from 'lucide-react';
+import type { ReactNode } from 'react';
+
+import { track } from '@/lib/analytics';
+import { cn } from '@/lib/utils';
+import type { OverviewConfigResult, OverviewModelSummary } from '@/lib/overview-data';
+import { buildOverviewDashboardHref } from '@/lib/overview-links';
+
+/**
+ * Model-level dashboard link. A plain anchor for the same reason the evidence
+ * links below are: the dashboard reads its filters from a snapshot `url-state.ts`
+ * takes at module evaluation, so a client-side `<Link>` navigation would land on
+ * an unfiltered dashboard while the label promises the model's own view.
+ */
+export function OverviewDetailLink({
+  href,
+  model,
+  ariaLabel,
+  className,
+  children,
+}: {
+  href: string;
+  model: string;
+  ariaLabel: string;
+  className?: string;
+  children: ReactNode;
+}) {
+  return (
+    <a
+      href={href}
+      aria-label={ariaLabel}
+      className={cn(
+        'group inline-flex min-h-11 items-center gap-1 whitespace-nowrap rounded-sm font-medium text-foreground underline decoration-brand/50 underline-offset-4 transition-colors hover:decoration-brand focus-visible:outline-none focus-visible:ring-[3px] focus-visible:ring-ring/50 motion-reduce:transition-none',
+        className,
+      )}
+      onClick={() => track('overview_model_detail_clicked', { model })}
+    >
+      {children}
+      <ArrowRight
+        aria-hidden="true"
+        className="size-4 transition-transform group-hover:translate-x-0.5 motion-reduce:transition-none"
+      />
+    </a>
+  );
+}
+
+const EVIDENCE_LINK_CLASS =
+  'group inline-flex min-h-11 w-fit items-center gap-1 whitespace-nowrap rounded-sm text-xs text-muted-foreground underline decoration-dotted underline-offset-4 transition-colors hover:text-foreground focus-visible:outline-none focus-visible:ring-[3px] focus-visible:ring-ring/50 motion-reduce:transition-none';
+
+/**
+ * One evidence link: the /inference dashboard pre-filtered to a ranked
+ * configuration. The visible text stays short so the compact rows scan; the
+ * full deployment topology rides in `ariaLabel`, which the caller also makes
+ * distinct per cohort — so no two evidence links share the accessible name
+ * "Open filtered dashboard".
+ *
+ * A plain anchor, not a `<Link>`: the dashboard reads its share-link params
+ * from a snapshot `url-state.ts` takes at module evaluation, so a client-side
+ * navigation would arrive with every filter dropped — silently unfiltered while
+ * the accessible name promises filtered evidence.
+ */
+export function OverviewDashboardLink({
+  locale,
+  model,
+  config,
+  ariaLabel,
+  children,
+}: {
+  locale: 'en' | 'zh';
+  model: OverviewModelSummary;
+  config: OverviewConfigResult;
+  ariaLabel: string;
+  children: ReactNode;
+}) {
+  return (
+    <a
+      href={buildOverviewDashboardHref(locale, model, config)}
+      aria-label={ariaLabel}
+      className={EVIDENCE_LINK_CLASS}
+    >
+      {children}
+    </a>
+  );
+}
diff --git a/packages/app/src/components/overview/overview-page.tsx b/packages/app/src/components/overview/overview-page.tsx
new file mode 100644
index 00000000..c99f3ae6
--- /dev/null
+++ b/packages/app/src/components/overview/overview-page.tsx
@@ -0,0 +1,59 @@
+import { Card } from '@/components/ui/card';
+import type { OverviewPageData } from '@/lib/overview-data';
+
+import {
+  DesktopOverviewMatrix,
+  MobileOverviewList,
+  OverviewMethodology,
+  overviewFormatters,
+  OVERVIEW_STRINGS,
+  type OverviewLocale,
+} from './overview-scorecard';
+
+interface OverviewPageProps {
+  data: OverviewPageData;
+  locale: OverviewLocale;
+}
+
+export function OverviewPageContent({ data, locale }: OverviewPageProps) {
+  const strings = OVERVIEW_STRINGS[locale];
+  const formatters = overviewFormatters(locale);
+  const snapshot =
+    data.datasetThroughDate === null
+      ? null
+      : strings.snapshot(formatters.shortDate(data.datasetThroughDate));
+
+  return (
+    <section className="pt-2 lg:pt-4">
+      <header className="max-w-4xl">
+        <h1 className="text-2xl lg:text-4xl font-bold tracking-tight">{strings.title}</h1>
+        <p className="mt-2 max-w-3xl text-sm text-muted-foreground sm:text-base lg:text-lg">
+          {strings.purpose}
+        </p>
+        <p className="mt-1.5 max-w-4xl text-xs leading-snug text-muted-foreground sm:text-sm">
+          {strings.scope}
+        </p>
+        {snapshot === null ? null : (
+          <p className="mt-1 text-xs text-muted-foreground/80 tabular-nums">{snapshot}</p>
+        )}
+      </header>
+
+      {/* Official-only summary; uploaded runs remain in the linked dashboard. */}
+      <Card className="mt-4 overflow-hidden p-0 md:p-0">
+        <DesktopOverviewMatrix
+          models={data.models}
+          locale={locale}
+          formatters={formatters}
+          strings={strings}
+        />
+        <MobileOverviewList
+          models={data.models}
+          locale={locale}
+          formatters={formatters}
+          strings={strings}
+        />
+        <OverviewMethodology strings={strings} />
+      </Card>
+    </section>
+  );
+}
diff --git a/packages/app/src/components/overview/overview-scorecard.tsx b/packages/app/src/components/overview/overview-scorecard.tsx
new file mode 100644
index 00000000..a1c02907
--- /dev/null
+++ b/packages/app/src/components/overview/overview-scorecard.tsx
@@ -0,0 +1,736 @@
+/**
+ * @file overview-scorecard.tsx
+ * @description The comparison surface itself. Desktop (a matrix at `xl`) and
+ * mobile (stacked cards below `xl`) render the SAME per-model hierarchy from one
+ * `ModelHierarchy` component, so the two surfaces cannot drift semantically.
+ *
+ * Each model reads top to bottom as: the primary precision's ranked hardware as
+ * horizontal chips (@50), the secondary precision (a subordinate ranked block or
+ * one compact coverage line), the @100 capability read, then a not-ranked line
+ * that accounts for every remaining hardware with an explicit reason. Precisions
+ * and incompatible cohorts are never compared across — each ranks only itself.
+ */
+
+import { Badge } from '@/components/ui/badge';
+import {
+  overviewHighValue,
+  OVERVIEW_HIGH_TIER,
+  OVERVIEW_PRIMARY_TIER,
+  type OverviewComparisonGroup,
+  type OverviewConfigResult,
+  type OverviewHardwareStatus,
+  type OverviewModelSummary,
+  type OverviewNotRankedEntry,
+} from '@/lib/overview-data';
+import { detailHref } from '@/lib/overview-links';
+
+import { OverviewDetailLink, OverviewDashboardLink } from './overview-detail-link';
+
+export type OverviewLocale = 'en' | 'zh';
+
+export const OVERVIEW_STRINGS = {
+  en: {
+    title: 'AI Inference Overview',
+    purpose: 'Compare validated serving results across active models and hardware.',
+    scope: `8K→1K · Single-turn · Output tok/s/GPU at ${OVERVIEW_PRIMARY_TIER} tok/s/user · Speculative decode only · FP4/FP8 by comparable coverage`,
+    snapshot: (through: string) => `Database snapshot through ${through}`,
+    modelHeader: 'Model',
+    resultsHeader: 'Ranked results',
+    detailsHeader: 'Details',
+    caption: `Validated output tok/s/GPU per model and comparable cohort at ${OVERVIEW_PRIMARY_TIER} and ${OVERVIEW_HIGH_TIER} tok/s/user.`,
+    cohort: (engine: string, mode: string) => `${engine} · ${mode}`,
+    aggregated: 'Aggregated',
+    disaggregated: 'Disaggregated',
+    anyEngine: 'All engines',
+    engineGroupLabels: {
+      trt: 'TRTLLM family',
+      sglang: 'SGLang/ATOM family',
+      vllm: 'vLLM family',
+    } as Record<string, string>,
+    engineGroupFallback: (engineGroup: string) => `${engineGroup.toUpperCase()} family`,
+    primaryHeading: (precision: string) => `${precision} · PRIMARY @${OVERVIEW_PRIMARY_TIER}`,
+    secondaryHeading: (precision: string) => `${precision} @${OVERVIEW_PRIMARY_TIER}`,
+    secondaryCoverage: (precision: string, list: string) =>
+      `${precision} coverage: ${list} measured; insufficient comparable results.`,
+    highHeading: `@${OVERVIEW_HIGH_TIER}`,
+    leader: 'Leader',
+    onlyExactResult: 'Only exact result',
+    highLeader: (hardware: string, value: string) => `${hardware} · ${value}`,
+    highLeaderSame: 'Same leader',
+    leaderChange: 'Leader change',
+    highInsufficient: `No exact read at ${OVERVIEW_HIGH_TIER} tok/s/user`,
+    highNoPrimaryBaseline: `No ${OVERVIEW_PRIMARY_TIER} tok/s/user baseline`,
+    evidenceHighDashboard: `Dashboard @${OVERVIEW_HIGH_TIER}`,
+    evidenceDashboardAria: (model: string, config: string) =>
+      `Open filtered dashboard: ${model} · ${config}`,
+    evidenceHighAria: (model: string, config: string) =>
+      `Open filtered dashboard @${OVERVIEW_HIGH_TIER} tok/s/user leader: ${model} · ${config}`,
+    notRankedLine: (entries: string) => `Not ranked: ${entries}`,
+    notRankedReasons: {
+      standard_decode_only: 'standard decode only',
+      int4_bf16_only: 'INT4/BF16 only',
+      different_serving_cohort: 'different serving cohort',
+      no_8k1k_data: 'no 8K/1K data',
+      cannot_reach_at50: `cannot reach @${OVERVIEW_PRIMARY_TIER}`,
+      no_exact_at50: `no exact @${OVERVIEW_PRIMARY_TIER} result`,
+    } as Record<string, string>,
+    otherPrecisionReason: (precisions: string[]) =>
+      `${precisions.map((precision) => precision.toUpperCase()).join('/')} only`,
+    joinList: (labels: string[]) =>
+      labels.length <= 1
+        ? (labels[0] ?? '')
+        : `${labels.slice(0, -1).join(', ')} and ${labels.at(-1) ?? ''}`,
+    methodologyNote:
+      'Precisions and incompatible serving cohorts are ranked separately. The default precision maximizes comparable hardware coverage. Not ranked does not mean slower.',
+    interpolationNote:
+      'Tier values interpolate each configuration’s official Pareto frontier — no extrapolation.',
+    cohortNote:
+      'Engine families with different speculative-decode acceptance forcing, and aggregated versus disaggregated deployments, rank as separate cohorts — never against each other.',
+    detailLink: 'View details',
+    detailAria: (modelLabel: string) => `View details: ${modelLabel}`,
+  },
+  zh: {
+    title: 'AI 推理总览',
+    purpose: '对比活跃模型在不同硬件上的已验证服务结果。',
+    scope: `8K→1K · 单轮 · ${OVERVIEW_PRIMARY_TIER} tok/s/user 下的输出 tok/s/GPU · 仅投机解码 · 按可比覆盖选择 FP4/FP8`,
+    snapshot: (through: string) => `数据库快照截至 ${through}`,
+    modelHeader: '模型',
+    resultsHeader: '排名结果',
+    detailsHeader: '详情',
+    caption: `各模型与可对比分组在 ${OVERVIEW_PRIMARY_TIER} 与 ${OVERVIEW_HIGH_TIER} tok/s/user 下的已验证输出 tok/s/GPU。`,
+    cohort: (engine: string, mode: string) => `${engine} · ${mode}`,
+    aggregated: '聚合部署',
+    disaggregated: '分离部署',
+    anyEngine: '全部引擎',
+    engineGroupLabels: {
+      trt: 'TRTLLM 系列',
+      sglang: 'SGLang/ATOM 系列',
+      vllm: 'vLLM 系列',
+    } as Record<string, string>,
+    engineGroupFallback: (engineGroup: string) => `${engineGroup.toUpperCase()} 系列`,
+    primaryHeading: (precision: string) => `${precision} · 主排名 @${OVERVIEW_PRIMARY_TIER}`,
+    secondaryHeading: (precision: string) => `${precision} @${OVERVIEW_PRIMARY_TIER}`,
+    secondaryCoverage: (precision: string, list: string) =>
+      `${precision} 覆盖：已测量 ${list}；可比结果不足。`,
+    highHeading: `@${OVERVIEW_HIGH_TIER}`,
+    leader: '领先',
+    onlyExactResult: '唯一精确读数',
+    highLeader: (hardware: string, value: string) => `${hardware} · ${value}`,
+    highLeaderSame: '领先者相同',
+    leaderChange: '领先者变化',
+    highInsufficient: `${OVERVIEW_HIGH_TIER} tok/s/user 下无精确读数`,
+    highNoPrimaryBaseline: `无 ${OVERVIEW_PRIMARY_TIER} tok/s/user 基线`,
+    evidenceHighDashboard: `仪表板 @${OVERVIEW_HIGH_TIER}`,
+    evidenceDashboardAria: (model: string, config: string) =>
+      `打开筛选后的仪表板：${model} · ${config}`,
+    evidenceHighAria: (model: string, config: string) =>
+      `打开筛选后的仪表板 @${OVERVIEW_HIGH_TIER} tok/s/user：${model} · ${config}`,
+    notRankedLine: (entries: string) => `未参与排名：${entries}`,
+    notRankedReasons: {
+      standard_decode_only: '仅标准解码',
+      int4_bf16_only: '仅 INT4/BF16',
+      different_serving_cohort: '不同服务配置',
+      no_8k1k_data: '无 8K/1K 数据',
+      cannot_reach_at50: `无法达到 @${OVERVIEW_PRIMARY_TIER}`,
+      no_exact_at50: `无精确 @${OVERVIEW_PRIMARY_TIER} 结果`,
+    } as Record<string, string>,
+    otherPrecisionReason: (precisions: string[]) =>
+      `仅 ${precisions.map((precision) => precision.toUpperCase()).join('/')}`,
+    joinList: (labels: string[]) => labels.join('、'),
+    methodologyNote:
+      '不同精度和不可直接比较的服务配置将分别排名。默认精度优先覆盖更多可比较硬件。未参与排名不代表性能更低。',
+    interpolationNote: '各档位数据基于各配置官方 Pareto 前沿插值；不进行外推。',
+    cohortNote:
+      '投机解码接受率强制方式不同的引擎系列，以及聚合与分离部署，按独立分组排名，彼此之间不作比较。',
+    detailLink: '查看详情',
+    detailAria: (modelLabel: string) => `查看详情：${modelLabel}`,
+  },
+} as const;
+
+export type OverviewStrings = (typeof OVERVIEW_STRINGS)[OverviewLocale];
+
+interface Formatters {
+  number: Intl.NumberFormat;
+  signed: Intl.NumberFormat;
+  date: (date: string) => string;
+  shortDate: (date: string) => string;
+}
+
+export function overviewFormatters(locale: OverviewLocale): Formatters {
+  const tag = locale === 'zh' ? 'zh-CN' : 'en-US';
+  const dateFormat = new Intl.DateTimeFormat(tag, {
+    year: 'numeric',
+    month: 'short',
+    day: 'numeric',
+    timeZone: 'UTC',
+  });
+  const shortDateFormat = new Intl.DateTimeFormat(tag, {
+    month: 'short',
+    day: 'numeric',
+    timeZone: 'UTC',
+  });
+  return {
+    number: new Intl.NumberFormat(tag, { maximumFractionDigits: 0 }),
+    signed: new Intl.NumberFormat(tag, { maximumFractionDigits: 0, signDisplay: 'exceptZero' }),
+    date: (date) => dateFormat.format(new Date(`${date}T00:00:00Z`)),
+    shortDate: (date) => shortDateFormat.format(new Date(`${date}T00:00:00Z`)),
+  };
+}
+
+/**
+ * A result's evidence dates, localized: a single day (`Jul 6` / `7月6日`) when
+ * both backing frontier points share a date, else an en-dash range
+ * (`Jun 24–Jul 4`). Null when the value has no backing date.
+ */
+function formatEvidenceDate(
+  formatters: Formatters,
+  evidenceDate: { from: string; to: string } | null,
+): string | null {
+  if (evidenceDate === null) return null;
+  const from = formatters.shortDate(evidenceDate.from);
+  return evidenceDate.from === evidenceDate.to
+    ? from
+    : `${from}–${formatters.shortDate(evidenceDate.to)}`;
+}
+
+function cohortLabel(
+  group: OverviewComparisonGroup,
+  strings: OverviewStrings,
+  showDbModel: boolean,
+): string {
+  const engine =
+    group.engineGroup === null
+      ? strings.anyEngine
+      : (strings.engineGroupLabels[group.engineGroup] ??
+        strings.engineGroupFallback(group.engineGroup));
+  const mode =
+    group.deploymentMode === 'disaggregated' ? strings.disaggregated : strings.aggregated;
+  const label = strings.cohort(engine, mode);
+  // Point releases of one display model rank in separate cohorts; name the raw
+  // db model so two same-engine, same-mode cohorts are told apart.
+  return showDbModel ? `${label} · ${group.dbModel}` : label;
+}
+
+/** Cohorts with at least one exact @50 read — the ones that render ranked chips. */
+function rankedCohorts(groups: readonly OverviewComparisonGroup[]): OverviewComparisonGroup[] {
+  return groups.filter((group) => group.primaryRanking.state !== 'insufficient_coverage');
+}
+
+/** A cohort's hardware that hold an exact @50 read, leader first then value desc. */
+function rankedStatuses(group: OverviewComparisonGroup): OverviewHardwareStatus[] {
+  return group.hardwareStatuses
+    .filter((status) => status.isPrimaryLeader || status.primaryDeltaPercent !== null)
+    .toSorted((a, b) => (b.primary.value ?? -1) - (a.primary.value ?? -1));
+}
+
+/** Not-ranked reason copy; `other_precision_only` names the measured precisions. */
+function notRankedReason(entry: OverviewNotRankedEntry, strings: OverviewStrings): string {
+  return entry.reason === 'other_precision_only'
+    ? strings.otherPrecisionReason(entry.precisions)
+    : strings.notRankedReasons[entry.reason];
+}
+
+/** The exact deployment topology of a ranked configuration, for an aria-label. */
+function evidenceConfigLabel(
+  config: OverviewConfigResult,
+  strings: OverviewStrings,
+  showDbModel: boolean,
+): string {
+  return [
+    config.hardwareLabel,
+    config.precision.toUpperCase(),
+    config.frameworkLabel,
+    config.specLabel,
+    config.disagg ? strings.disaggregated : strings.aggregated,
+    config.parallelism,
+    `${config.totalGpu} GPU`,
+    ...(showDbModel ? [config.dbModel] : []),
+  ].join(' · ');
+}
+
+const BLOCK_HEADING_CLASS =
+  'mb-0.5 text-xs font-semibold uppercase tracking-wider text-muted-foreground';
+const SUBORDINATE_HEADING_CLASS =
+  'mb-0.5 text-[11px] font-medium uppercase tracking-wider text-muted-foreground/80';
+
+/**
+ * One ranked hardware: its label, its @50 value linking to the filtered
+ * dashboard, and its standing — `Leader`, its signed delta against the leader,
+ * or `Only exact result` when it is the cohort's lone exact read.
+ */
+function RankedChip({
+  locale,
+  model,
+  group,
+  status,
+  multiDbModel,
+  formatters,
+  strings,
+}: {
+  locale: OverviewLocale;
+  model: OverviewModelSummary;
+  group: OverviewComparisonGroup;
+  status: OverviewHardwareStatus;
+  multiDbModel: boolean;
+  formatters: Formatters;
+  strings: OverviewStrings;
+}) {
+  const { value, config } = status.primary;
+  const standing = status.isPrimaryLeader
+    ? group.primaryRanking.state === 'single_measured'
+      ? strings.onlyExactResult
+      : strings.leader
+    : `${formatters.signed.format(status.primaryDeltaPercent as number)}%`;
+  const date = formatEvidenceDate(formatters, status.primary.evidenceDate);
+  return (
+    <p className="flex flex-wrap items-baseline gap-x-1.5 gap-y-0.5">
+      <span className="font-medium">{status.hardwareLabel}</span>
+      {value === null ? null : config === null ? (
+        <span className="text-sm font-semibold tabular-nums">
+          {formatters.number.format(value)}
+        </span>
+      ) : (
+        <OverviewDashboardLink
+          locale={locale}
+          model={model}
+          config={config}
+          ariaLabel={strings.evidenceDashboardAria(
+            model.modelLabel,
+            evidenceConfigLabel(config, strings, multiDbModel),
+          )}
+        >
+          <span className="text-sm font-semibold tabular-nums text-foreground">
+            {formatters.number.format(value)}
+          </span>
+        </OverviewDashboardLink>
+      )}
+      <span className="text-xs text-muted-foreground tabular-nums">· {standing}</span>
+      {date === null ? null : (
+        <span className="text-xs text-muted-foreground/80 tabular-nums">· {date}</span>
+      )}
+    </p>
+  );
+}
+
+/** A precision's ranked cohorts as stacked chips, under one heading. */
+function RankedSection({
+  locale,
+  model,
+  cohorts,
+  heading,
+  headingClass,
+  formatters,
+  strings,
+}: {
+  locale: OverviewLocale;
+  model: OverviewModelSummary;
+  cohorts: OverviewComparisonGroup[];
+  heading: string;
+  headingClass: string;
+  formatters: Formatters;
+  strings: OverviewStrings;
+}) {
+  if (cohorts.length === 0) return null;
+  const multiDbModel = new Set(cohorts.map((group) => group.dbModel)).size > 1;
+  return (
+    <div>
+      <p className={headingClass}>{heading}</p>
+      {cohorts.map((group) => (
+        <div key={group.id} className="space-y-0.5">
+          {cohorts.length > 1 ? (
+            <p className="text-xs text-muted-foreground">
+              {cohortLabel(group, strings, multiDbModel)}
+            </p>
+          ) : null}
+          {rankedStatuses(group).map((status) => (
+            <RankedChip
+              key={status.hardware}
+              locale={locale}
+              model={model}
+              group={group}
+              status={status}
+              multiDbModel={multiDbModel}
+              formatters={formatters}
+              strings={strings}
+            />
+          ))}
+        </div>
+      ))}
+    </div>
+  );
+}
+
+/** The primary precision block: its heading names the precision and the tier. */
+function PrimarySection({
+  locale,
+  model,
+  cohorts,
+  formatters,
+  strings,
+}: {
+  locale: OverviewLocale;
+  model: OverviewModelSummary;
+  cohorts: OverviewComparisonGroup[];
+  formatters: Formatters;
+  strings: OverviewStrings;
+}) {
+  const precision = (model.selectedPrecision ?? '').toUpperCase();
+  return (
+    <RankedSection
+      locale={locale}
+      model={model}
+      cohorts={cohorts}
+      heading={strings.primaryHeading(precision)}
+      headingClass={BLOCK_HEADING_CLASS}
+      formatters={formatters}
+      strings={strings}
+    />
+  );
+}
+
+/**
+ * The secondary precision, subordinate to the primary: a ranked block when it
+ * genuinely adds comparable hardware, otherwise one compact coverage line so the
+ * precision still surfaces without a second full table.
+ */
+function SecondarySection({
+  locale,
+  model,
+  formatters,
+  strings,
+}: {
+  locale: OverviewLocale;
+  model: OverviewModelSummary;
+  formatters: Formatters;
+  strings: OverviewStrings;
+}) {
+  const secondary = model.secondary;
+  if (secondary === null) return null;
+  const precision = secondary.precision.toUpperCase();
+  if (secondary.state === 'coverage') {
+    const list = strings.joinList(secondary.measuredHardware);
+    if (list === '') return null;
+    return (
+      <p className="text-xs text-muted-foreground">{strings.secondaryCoverage(precision, list)}</p>
+    );
+  }
+  return (
+    <RankedSection
+      locale={locale}
+      model={model}
+      cohorts={rankedCohorts(secondary.comparisonGroups)}
+      heading={strings.secondaryHeading(precision)}
+      headingClass={SUBORDINATE_HEADING_CLASS}
+      formatters={formatters}
+      strings={strings}
+    />
+  );
+}
+
+/**
+ * The 100 tok/s/user read of one cohort. It is ranked from its own evidence, so
+ * it names its own leader and says plainly whether that leader changed.
+ */
+function HighTierRead({
+  group,
+  formatters,
+  strings,
+}: {
+  group: OverviewComparisonGroup;
+  formatters: Formatters;
+  strings: OverviewStrings;
+}) {
+  const leader = group.highRanking.leader;
+  const value = leader === null ? null : overviewHighValue(leader);
+  if (leader === null || value === null) {
+    return <span className="text-xs text-muted-foreground">{strings.highInsufficient}</span>;
+  }
+  // The @100 leader's own read date — from its hardware's high read, so a @100
+  // read backed by a different config than the @50 leader shows its own date.
+  const date = formatEvidenceDate(
+    formatters,
+    group.hardwareStatuses.find((status) => status.hardware === leader.hardware)?.high
+      .evidenceDate ?? null,
+  );
+  return (
+    <span className="flex flex-wrap items-baseline gap-x-2 gap-y-0.5">
+      <span className="tabular-nums">
+        {strings.highLeader(leader.hardwareLabel, formatters.number.format(value))}
+      </span>
+      {group.highRanking.state === 'single_measured' ? (
+        <span className="text-xs text-muted-foreground">{strings.onlyExactResult}</span>
+      ) : group.highLeaderTransition === 'changed_hardware' ? (
+        <Badge
+          variant="outline"
+          className="border-amber-500/40 bg-amber-500/10 text-amber-700 dark:border-amber-400/30 dark:text-amber-400"
+          title={strings.leaderChange}
+        >
+          {strings.leaderChange}
+        </Badge>
+      ) : group.highLeaderTransition === 'no_primary_baseline' ? (
+        <span className="text-xs text-muted-foreground">{strings.highNoPrimaryBaseline}</span>
+      ) : (
+        <span className="text-xs text-muted-foreground">{strings.highLeaderSame}</span>
+      )}
+      {date === null ? null : (
+        <span className="text-xs text-muted-foreground/80 tabular-nums">· {date}</span>
+      )}
+    </span>
+  );
+}
+
+/**
+ * The @100 capability read per comparable cohort. When the @100 leader runs a
+ * different configuration than the @50 leader, its own filtered dashboard rides
+ * along so the high-tier read ships with the evidence behind it.
+ */
+function HighSection({
+  locale,
+  model,
+  cohorts,
+  formatters,
+  strings,
+}: {
+  locale: OverviewLocale;
+  model: OverviewModelSummary;
+  cohorts: OverviewComparisonGroup[];
+  formatters: Formatters;
+  strings: OverviewStrings;
+}) {
+  const highCohorts = cohorts.filter((group) => group.primaryRanking.state === 'comparable');
+  if (highCohorts.length === 0) return null;
+  const multiDbModel = new Set(highCohorts.map((group) => group.dbModel)).size > 1;
+  return (
+    <div>
+      <p className={BLOCK_HEADING_CLASS}>{strings.highHeading}</p>
+      {highCohorts.map((group) => {
+        const primaryLeader = group.primaryRanking.leader;
+        const highLeader = group.highRanking.leader;
+        return (
+          <div key={group.id} className="flex flex-wrap items-baseline gap-x-3 gap-y-0.5 text-sm">
+            {highCohorts.length > 1 ? (
+              <span className="text-xs text-muted-foreground">
+                {cohortLabel(group, strings, multiDbModel)}
+              </span>
+            ) : null}
+            <HighTierRead group={group} formatters={formatters} strings={strings} />
+            {highLeader !== null &&
+            (primaryLeader === null || highLeader.key !== primaryLeader.key) ? (
+              <OverviewDashboardLink
+                locale={locale}
+                model={model}
+                config={highLeader}
+                ariaLabel={strings.evidenceHighAria(
+                  model.modelLabel,
+                  evidenceConfigLabel(highLeader, strings, multiDbModel),
+                )}
+              >
+                {strings.evidenceHighDashboard}
+              </OverviewDashboardLink>
+            ) : null}
+          </div>
+        );
+      })}
+    </div>
+  );
+}
+
+/**
+ * Every remaining hardware, accounted for in one line: those with no exact @50
+ * read and not ranked by the secondary block, each with its explicit reason. The
+ * standing note "not ranked does not mean slower" lives once in the methodology.
+ */
+function NotRankedLine({
+  model,
+  strings,
+}: {
+  model: OverviewModelSummary;
+  strings: OverviewStrings;
+}) {
+  // Hardware already named by the secondary section — its ranked rows or its
+  // coverage 'measured' list — is not repeated here; dedup against the secondary's
+  // measured labels so each hardware surfaces in exactly one place.
+  const secondaryMeasured = new Set(model.secondary?.measuredHardware);
+  const entries = model.notRanked
+    .filter((entry) => !secondaryMeasured.has(entry.hardwareLabel))
+    .map((entry) => `${entry.hardwareLabel} — ${notRankedReason(entry, strings)}`)
+    .join(' · ');
+  if (entries === '') return null;
+  return <p className="text-xs text-muted-foreground">{strings.notRankedLine(entries)}</p>;
+}
+
+/** The full per-model hierarchy, identical on both surfaces. */
+function ModelHierarchy({
+  locale,
+  model,
+  formatters,
+  strings,
+}: {
+  locale: OverviewLocale;
+  model: OverviewModelSummary;
+  formatters: Formatters;
+  strings: OverviewStrings;
+}) {
+  const cohorts = rankedCohorts(model.comparisonGroups);
+  return (
+    <div className="space-y-2 text-sm">
+      <PrimarySection
+        locale={locale}
+        model={model}
+        cohorts={cohorts}
+        formatters={formatters}
+        strings={strings}
+      />
+      <SecondarySection locale={locale} model={model} formatters={formatters} strings={strings} />
+      <HighSection
+        locale={locale}
+        model={model}
+        cohorts={cohorts}
+        formatters={formatters}
+        strings={strings}
+      />
+      <NotRankedLine model={model} strings={strings} />
+    </div>
+  );
+}
+
+/**
+ * The model's name and — for a coverage-only model that shows no ranked chips —
+ * its freshness. A ranked model dates each result on its own chip instead, so
+ * the model-level date is only the fallback for a model with nothing to rank.
+ */
+function ModelIdentity({
+  model,
+  formatters,
+}: {
+  model: OverviewModelSummary;
+  formatters: Formatters;
+}) {
+  const hasRankedResults =
+    rankedCohorts(model.comparisonGroups).length > 0 || model.secondary?.state === 'ranked';
+  const freshness =
+    hasRankedResults || model.latestWorkloadDate === null
+      ? null
+      : formatters.date(model.latestWorkloadDate);
+  return (
+    <div className="flex min-w-0 flex-wrap items-baseline gap-x-2 gap-y-1">
+      <h2 className="text-sm font-semibold leading-snug">{model.modelLabel}</h2>
+      {freshness ? (
+        <span className="text-xs text-muted-foreground tabular-nums">{freshness}</span>
+      ) : null}
+    </div>
+  );
+}
+
+interface SurfaceProps {
+  models: OverviewModelSummary[];
+  locale: OverviewLocale;
+  formatters: Formatters;
+  strings: OverviewStrings;
+}
+
+/** Model column plus one content column carrying the full hierarchy; no scroll. */
+export function DesktopOverviewMatrix({ models, locale, formatters, strings }: SurfaceProps) {
+  return (
+    <div className="hidden xl:block">
+      <table
+        data-testid="overview-desktop-matrix"
+        className="w-full table-fixed border-collapse text-sm"
+      >
+        <caption className="sr-only">{strings.caption}</caption>
+        <colgroup>
+          <col className="w-[18%]" />
+          <col className="w-[70%]" />
+          <col className="w-[12%]" />
+        </colgroup>
+        <thead>
+          <tr className="border-b border-border/50 text-xs uppercase tracking-wider text-muted-foreground">
+            <th scope="col" className="px-4 py-2 text-left font-semibold lg:px-6">
+              {strings.modelHeader}
+            </th>
+            <th scope="col" className="px-4 py-2 text-left font-semibold">
+              {strings.resultsHeader}
+            </th>
+            <th scope="col" className="px-4 py-2 text-left font-semibold">
+              {strings.detailsHeader}
+            </th>
+          </tr>
+        </thead>
+        {models.map((model) => (
+          <tbody
+            key={model.model}
+            data-testid="overview-desktop-model"
+            data-model={model.model}
+            className="border-b border-border/50 last:border-b-0"
+          >
+            <tr className="align-top">
+              <th scope="row" className="px-4 py-3 text-left align-top font-normal lg:px-6">
+                <ModelIdentity model={model} formatters={formatters} />
+              </th>
+              <td className="px-4 py-3">
+                <ModelHierarchy
+                  locale={locale}
+                  model={model}
+                  formatters={formatters}
+                  strings={strings}
+                />
+              </td>
+              <td className="px-4 py-3 align-top">
+                <OverviewDetailLink
+                  href={detailHref(locale, model)}
+                  model={model.model}
+                  ariaLabel={strings.detailAria(model.modelLabel)}
+                >
+                  {strings.detailLink}
+                </OverviewDetailLink>
+              </td>
+            </tr>
+          </tbody>
+        ))}
+      </table>
+    </div>
+  );
+}
+
+/** Below `xl`: the same hierarchy stacked as cards, always fully visible. */
+export function MobileOverviewList({ models, locale, formatters, strings }: SurfaceProps) {
+  return (
+    <ul data-testid="overview-mobile-list" className="divide-y divide-border/50 xl:hidden">
+      {models.map((model) => (
+        <li key={model.model}>
+          <article
+            data-testid="overview-mobile-model"
+            data-model={model.model}
+            className="space-y-2 px-4 py-3"
+          >
+            <ModelIdentity model={model} formatters={formatters} />
+            <ModelHierarchy
+              locale={locale}
+              model={model}
+              formatters={formatters}
+              strings={strings}
+            />
+            <OverviewDetailLink
+              href={detailHref(locale, model)}
+              model={model.model}
+              ariaLabel={strings.detailAria(model.modelLabel)}
+              className="min-h-11 w-full justify-between"
+            >
+              {strings.detailLink}
+            </OverviewDetailLink>
+          </article>
+        </li>
+      ))}
+    </ul>
+  );
+}
+
+export function OverviewMethodology({ strings }: { strings: OverviewStrings }) {
+  return (
+    <div className="space-y-1 border-t border-border/50 px-4 py-3 text-xs leading-snug text-muted-foreground lg:px-6">
+      <p>{strings.methodologyNote}</p>
+      <p>{strings.cohortNote}</p>
+      <p>{strings.interpolationNote}</p>
+    </div>
+  );
+}
diff --git a/packages/app/src/components/ui/mode-toggle.tsx b/packages/app/src/components/ui/mode-toggle.tsx
index f216624f..7dcdd354 100644
--- a/packages/app/src/components/ui/mode-toggle.tsx
+++ b/packages/app/src/components/ui/mode-toggle.tsx
@@ -27,7 +27,7 @@ export function ModeToggle() {
   };
 
   const buttonClasses = cn(
-    'inline-flex items-center justify-center rounded-md p-2',
+    'inline-flex items-center justify-center rounded-md size-11',
     'text-muted-foreground hover:text-foreground hover:bg-accent',
     'transition-colors duration-200',
     'focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-ring focus-visible:ring-offset-2',
diff --git a/packages/app/src/hooks/useTogglableSet.ts b/packages/app/src/hooks/useTogglableSet.ts
index bf84694c..59bbac71 100644
--- a/packages/app/src/hooks/useTogglableSet.ts
+++ b/packages/app/src/hooks/useTogglableSet.ts
@@ -1,33 +1,8 @@
 import { useCallback, useState } from 'react';
 
-/**
- * Pure toggle state transition.
- *
- * - If all items are active and one is toggled, only that item remains active (solo).
- * - If only one item is active and it's toggled, all items become active (restore all).
- * - Otherwise, the item is simply added or removed from the active set.
- */
-export function computeToggle(prev: Set<string>, item: string, allItems: Set<string>): Set<string> {
-  const allAreActive = prev.size === allItems.size;
-  const isActive = prev.has(item);
+import { computeToggle } from '@/lib/toggle-set';
 
-  if (isActive) {
-    if (allAreActive) {
-      // Solo: only keep the clicked item
-      return new Set([item]);
-    } else if (prev.size === 1) {
-      // Restore all: re-enable everything
-      return allItems;
-    }
-    // Remove the clicked item
-    const next = new Set(prev);
-    next.delete(item);
-    return next;
-  }
-  // Add the clicked item
-  const next = new Set([...prev, item]);
-  return next;
-}
+export { computeToggle };
 
 /**
  * Hook for managing a togglable set with "click to solo, click again to restore all" behavior.
diff --git a/packages/app/src/lib/benchmark-data.server.ts b/packages/app/src/lib/benchmark-data.server.ts
new file mode 100644
index 00000000..ccc85317
--- /dev/null
+++ b/packages/app/src/lib/benchmark-data.server.ts
@@ -0,0 +1,21 @@
+import { FIXTURES_MODE, getDb } from '@semianalysisai/inferencex-db/connection';
+import {
+  type BenchmarkRow,
+  getLatestBenchmarks,
+} from '@semianalysisai/inferencex-db/queries/benchmarks';
+
+import { cachedQuery } from '@/lib/api-cache';
+import { loadFixture } from '@/lib/test-fixtures';
+
+/** Cache slot is keyed on the dbKeys array. Both `/compare/<slug>` and
+ *  `/compare-per-dollar/<slug>` for the same model hit the same blob entry —
+ *  the per-dollar route doesn't duplicate the fetch or the cache. */
+export const getCachedBenchmarks = cachedQuery(
+  (dbModelKeys: string[]) => {
+    if (FIXTURES_MODE) return Promise.resolve(loadFixture<BenchmarkRow[]>('benchmarks'));
+
+    return getLatestBenchmarks(getDb(), dbModelKeys);
+  },
+  'benchmarks',
+  { blobOnly: true },
+);
diff --git a/packages/app/src/lib/compare-ssr.ts b/packages/app/src/lib/compare-ssr.ts
index 592a8238..5a3ec2c6 100644
--- a/packages/app/src/lib/compare-ssr.ts
+++ b/packages/app/src/lib/compare-ssr.ts
@@ -18,16 +18,11 @@ import {
   SITE_URL,
   sequenceToIslOsl,
 } from '@semianalysisai/inferencex-constants';
-import { FIXTURES_MODE, getDb } from '@semianalysisai/inferencex-db/connection';
-
-import {
-  type BenchmarkRow,
-  getLatestBenchmarks,
-} from '@semianalysisai/inferencex-db/queries/benchmarks';
+import type { BenchmarkRow } from '@semianalysisai/inferencex-db/queries/benchmarks';
 
 import { interpolateForGPU } from '@/components/calculator/interpolation';
 import type { GPUDataPoint, InterpolatedResult } from '@/components/calculator/types';
-import { cachedQuery } from '@/lib/api-cache';
+import { getCachedBenchmarks } from '@/lib/benchmark-data.server';
 import { rowToAggDataEntry } from '@/lib/benchmark-transform';
 import { getHardwareKey } from '@/lib/chart-utils';
 import {
@@ -39,24 +34,8 @@ import {
   compareModelSeoName,
 } from '@/lib/compare-slug';
 import { getHardwareConfig, getGpuSpecs } from '@/lib/constants';
-import { loadFixture } from '@/lib/test-fixtures';
-
-// ---------------------------------------------------------------------------
-// Cached benchmark fetch
-// ---------------------------------------------------------------------------
 
-/** Cache slot is keyed on the dbKeys array. Both `/compare/<slug>` and
- *  `/compare-per-dollar/<slug>` for the same model hit the same blob entry —
- *  the per-dollar route doesn't duplicate the fetch or the cache. */
-export const getCachedBenchmarks = cachedQuery(
-  (dbModelKeys: string[]) => {
-    if (FIXTURES_MODE) return Promise.resolve(loadFixture<BenchmarkRow[]>('benchmarks'));
-
-    return getLatestBenchmarks(getDb(), dbModelKeys);
-  },
-  'benchmarks',
-  { blobOnly: true },
-);
+export { getCachedBenchmarks };
 
 // ---------------------------------------------------------------------------
 // URL-param validators (shared by both routes' overrides)
diff --git a/packages/app/src/lib/exclusion.ts b/packages/app/src/lib/exclusion.ts
index 21f4d1bd..941202f1 100644
--- a/packages/app/src/lib/exclusion.ts
+++ b/packages/app/src/lib/exclusion.ts
@@ -1,6 +1,6 @@
 import { SPEC_METHOD_KEYS } from '@semianalysisai/inferencex-constants';
 
-import { computeToggle } from '@/hooks/useTogglableSet';
+import { computeToggle } from '@/lib/toggle-set';
 
 /**
  * Data-driven config exclusion.
diff --git a/packages/app/src/lib/i18n.test.ts b/packages/app/src/lib/i18n.test.ts
index 492d2742..36a9fde3 100644
--- a/packages/app/src/lib/i18n.test.ts
+++ b/packages/app/src/lib/i18n.test.ts
@@ -40,6 +40,7 @@ describe('hasZhSibling', () => {
   it('matches mirrored exact routes', () => {
     expect(hasZhSibling('/')).toBe(true);
     expect(hasZhSibling('/inference')).toBe(true);
+    expect(hasZhSibling('/overview')).toBe(true);
     expect(hasZhSibling('/about')).toBe(true);
   });
 
@@ -72,12 +73,14 @@ describe('switchLocalePath', () => {
   it('switches English pages to their zh sibling', () => {
     expect(switchLocalePath('/')).toBe('/zh');
     expect(switchLocalePath('/inference')).toBe('/zh/inference');
+    expect(switchLocalePath('/overview')).toBe('/zh/overview');
     expect(switchLocalePath('/blog/some-post')).toBe('/zh/blog/some-post');
   });
 
   it('switches zh pages back to English', () => {
     expect(switchLocalePath('/zh')).toBe('/');
     expect(switchLocalePath('/zh/quotes')).toBe('/quotes');
+    expect(switchLocalePath('/zh/overview')).toBe('/overview');
     expect(switchLocalePath('/zh/blog/some-post')).toBe('/blog/some-post');
   });
 
diff --git a/packages/app/src/lib/i18n.ts b/packages/app/src/lib/i18n.ts
index b094b456..7577353b 100644
--- a/packages/app/src/lib/i18n.ts
+++ b/packages/app/src/lib/i18n.ts
@@ -36,6 +36,7 @@ export function isZhPathname(pathname: string): boolean {
  */
 export const ZH_MIRRORED_ROUTES: readonly { path: string; exact?: boolean }[] = [
   { path: '/', exact: true },
+  { path: '/overview', exact: true },
   { path: '/inference', exact: true },
   { path: '/inference/agentic' },
   { path: '/evaluation', exact: true },
diff --git a/packages/app/src/lib/overview-config-identity.test.ts b/packages/app/src/lib/overview-config-identity.test.ts
new file mode 100644
index 00000000..02fab087
--- /dev/null
+++ b/packages/app/src/lib/overview-config-identity.test.ts
@@ -0,0 +1,97 @@
+import { describe, expect, it } from 'vitest';
+
+import type { BenchmarkRow } from './api';
+import { overviewConfigIdentityKey } from './overview-config-identity';
+
+let nextId = 1;
+
+function row(overrides: Partial<BenchmarkRow> = {}): BenchmarkRow {
+  return {
+    id: nextId++,
+    hardware: 'b200',
+    framework: 'sglang',
+    model: 'qwen3.5',
+    precision: 'fp8',
+    spec_method: 'mtp',
+    disagg: false,
+    is_multinode: false,
+    prefill_tp: 8,
+    prefill_ep: 1,
+    prefill_dp_attention: false,
+    prefill_num_workers: 1,
+    decode_tp: 8,
+    decode_ep: 1,
+    decode_dp_attention: false,
+    decode_num_workers: 1,
+    num_prefill_gpu: 8,
+    // Not 8: the num_decode_gpu case below flips this to 8 and must change the key.
+    num_decode_gpu: 4,
+    benchmark_type: 'single_turn',
+    isl: 8192,
+    osl: 1024,
+    conc: 16,
+    offload_mode: 'off',
+    image: null,
+    metrics: { median_intvty: 50, output_tput_per_gpu: 1000 },
+    date: '2026-07-20',
+    run_url: null,
+    ...overrides,
+  };
+}
+
+/** The pre-v2.1 coarse key: hardware/framework/spec/disagg/offload, no GPU counts. */
+const oldCoarseKey = (r: BenchmarkRow): string =>
+  `${r.hardware}|${r.framework}|${r.spec_method}|${r.disagg}|${r.offload_mode}`;
+
+describe('overviewConfigIdentityKey', () => {
+  it.each([
+    ['model', 'kimik2.7-code'],
+    ['hardware', 'gb300'],
+    ['framework', 'dynamo-trt'],
+    ['precision', 'fp4'],
+    ['spec_method', 'eagle'],
+    ['disagg', true],
+    ['is_multinode', true],
+    ['prefill_tp', 4],
+    ['prefill_ep', 4],
+    ['prefill_dp_attention', true],
+    ['prefill_num_workers', 7],
+    ['decode_tp', 16],
+    ['decode_ep', 16],
+    ['decode_dp_attention', true],
+    ['decode_num_workers', 2],
+    ['num_prefill_gpu', 28],
+    ['num_decode_gpu', 8],
+    ['offload_mode', 'cpu'],
+  ] satisfies [keyof BenchmarkRow, BenchmarkRow[keyof BenchmarkRow]][])(
+    'changes when %s changes',
+    (field, value) => {
+      expect(overviewConfigIdentityKey(row({ [field]: value } as Partial<BenchmarkRow>))).not.toBe(
+        overviewConfigIdentityKey(row()),
+      );
+    },
+  );
+
+  it('keeps concurrency and evidence outside configuration identity', () => {
+    const base = row();
+    expect(
+      overviewConfigIdentityKey({
+        ...base,
+        conc: 64,
+        date: '2026-07-21',
+        image: 'new-image',
+        run_url: 'https://github.com/SemiAnalysisAI/InferenceX/actions/runs/2',
+      }),
+    ).toBe(overviewConfigIdentityKey(base));
+  });
+
+  it('separates disagg topologies the old coarse key merged (28P+8D vs 40P+16D)', () => {
+    // The old coarse key omitted GPU counts, so two different P/D splits on the
+    // same hardware/framework/spec/disagg/offload collided into one result.
+    const split28p8d = row({ disagg: true, num_prefill_gpu: 28, num_decode_gpu: 8 });
+    const split40p16d = row({ disagg: true, num_prefill_gpu: 40, num_decode_gpu: 16 });
+
+    expect(oldCoarseKey(split28p8d)).toBe(oldCoarseKey(split40p16d));
+    expect(overviewConfigIdentityKey(split28p8d)).not.toBe(overviewConfigIdentityKey(split40p16d));
+  });
+});
diff --git a/packages/app/src/lib/overview-config-identity.ts b/packages/app/src/lib/overview-config-identity.ts
new file mode 100644
index 00000000..18e4e9f4
--- /dev/null
+++ b/packages/app/src/lib/overview-config-identity.ts
@@ -0,0 +1,33 @@
+import type { BenchmarkRow } from './api';
+
+/**
+ * Exact per-deployment identity for the overview grouping. Every dimension that
+ * makes a serving configuration a distinct deployable topology is part of the
+ * key — model, hardware, framework, precision, spec method, disagg, multinode,
+ * per-role parallelism and worker counts, GPU counts, and offload mode. Rows
+ * that differ only in concurrency, date, image, or run URL share one identity.
+ *
+ * Serialized as a JSON tuple so no field delimiter can collide with a value.
+ */
+export function overviewConfigIdentityKey(row: BenchmarkRow): string {
+  return JSON.stringify([
+    row.model,
+    row.hardware,
+    row.framework,
+    row.precision,
+    row.spec_method,
+    row.disagg,
+    row.is_multinode,
+    row.prefill_tp,
+    row.prefill_ep,
+    row.prefill_dp_attention,
+    row.prefill_num_workers,
+    row.decode_tp,
+    row.decode_ep,
+    row.decode_dp_attention,
+    row.decode_num_workers,
+    row.num_prefill_gpu,
+    row.num_decode_gpu,
+    row.offload_mode ?? 'off',
+  ]);
+}
diff --git a/packages/app/src/lib/overview-data.server.ts b/packages/app/src/lib/overview-data.server.ts
new file mode 100644
index 00000000..a9d3cafa
--- /dev/null
+++ b/packages/app/src/lib/overview-data.server.ts
@@ -0,0 +1,29 @@
+import { DISPLAY_MODEL_TO_DB } from '@semianalysisai/inferencex-constants';
+import { FIXTURES_MODE } from '@semianalysisai/inferencex-db/connection';
+
+import type { BenchmarkRow } from '@/lib/api';
+import { getCachedBenchmarks } from '@/lib/benchmark-data.server';
+import { DEFAULT_MODELS } from '@/lib/data-mappings';
+import { assembleOverviewPageData, type OverviewPageData } from '@/lib/overview-data';
+import { loadFixture } from '@/lib/test-fixtures';
+
+export async function getOverviewPageData(): Promise<OverviewPageData> {
+  // E2E fixtures mode serves a small synthetic rows-by-display-model fixture
+  // through the same assembler the live path uses, so a contract drift breaks
+  // the fixture tests instead of silently stranding the page's data.
+  if (FIXTURES_MODE) {
+    return assembleOverviewPageData(loadFixture<Record<string, BenchmarkRow[]>>('overview-rows'));
+  }
+
+  // Fetch rows per db model, concatenated per display model (one display model
+  // can span several db buckets), then hand the same shape to the assembler.
+  const entries = await Promise.all(
+    [...DEFAULT_MODELS].map(async (model) => {
+      const keys = DISPLAY_MODEL_TO_DB[model] ?? [];
+      const rows = keys.length > 0 ? await getCachedBenchmarks(keys) : [];
+      return [model, rows] as const;
+    }),
+  );
+
+  return assembleOverviewPageData(Object.fromEntries(entries));
+}
diff --git a/packages/app/src/lib/overview-data.test.ts b/packages/app/src/lib/overview-data.test.ts
new file mode 100644
index 00000000..96df49f9
--- /dev/null
+++ b/packages/app/src/lib/overview-data.test.ts
@@ -0,0 +1,740 @@
+import { describe, expect, it } from 'vitest';
+
+import overviewRowsFixture from '../../cypress/fixtures/api/overview-rows.json';
+
+import type { BenchmarkRow } from './api';
+import { getHardwareConfig } from './constants';
+import { DEFAULT_MODELS, Model, Precision } from './data-mappings';
+import { overviewConfigIdentityKey } from './overview-config-identity';
+import {
+  assembleOverviewPageData,
+  buildOverviewHardwareOrder,
+  buildOverviewModelSummary,
+  overviewPrimaryValue,
+  selectOverviewPrecision,
+  type OverviewModelSummary,
+} from './overview-data';
+
+let nextId = 1;
+
+function row(overrides: Partial<BenchmarkRow> = {}): BenchmarkRow {
+  return {
+    id: nextId++,
+    hardware: 'b200',
+    framework: 'sglang',
+    model: 'qwen3.5',
+    precision: 'fp8',
+    spec_method: 'mtp',
+    disagg: false,
+    is_multinode: false,
+    prefill_tp: 8,
+    prefill_ep: 1,
+    prefill_dp_attention: false,
+    prefill_num_workers: 1,
+    decode_tp: 8,
+    decode_ep: 1,
+    decode_dp_attention: false,
+    decode_num_workers: 1,
+    num_prefill_gpu: 8,
+    num_decode_gpu: 8,
+    benchmark_type: 'single_turn',
+    isl: 8192,
+    osl: 1024,
+    conc: 16,
+    offload_mode: 'off',
+    image: null,
+    metrics: { median_intvty: 50, output_tput_per_gpu: 1000 },
+    date: '2026-07-20',
+    run_url: null,
+    ...overrides,
+  };
+}
+
+/** One frontier point per tier for a single configuration. */
+function frontier(
+  throughputs: [number, number, number, number],
+  overrides: Partial<BenchmarkRow> = {},
+): BenchmarkRow[] {
+  return [30, 50, 75, 100].map((tier, index) =>
+    row({
+      conc: index + 1,
+      metrics: { median_intvty: tier, output_tput_per_gpu: throughputs[index] },
+      ...overrides,
+    }),
+  );
+}
+
+/** Frontier at explicit [interactivity, throughput] knots — for clamped/unreachable tiers. */
+function frontierAt(
+  points: [number, number][],
+  overrides: Partial<BenchmarkRow> = {},
+): BenchmarkRow[] {
+  return points.map(([intvty, tput], index) =>
+    row({
+      conc: index + 1,
+      metrics: { median_intvty: intvty, output_tput_per_gpu: tput },
+      ...overrides,
+    }),
+  );
+}
+
+/** One config's frontier at explicit [interactivity, throughput, date] knots. */
+function datedFrontier(
+  knots: [number, number, string][],
+  overrides: Partial<BenchmarkRow> = {},
+): BenchmarkRow[] {
+  return knots.map(([intvty, tput, date], index) =>
+    row({
+      conc: index + 1,
+      metrics: { median_intvty: intvty, output_tput_per_gpu: tput },
+      date,
+      ...overrides,
+    }),
+  );
+}
+
+function statusOf(summary: OverviewModelSummary, hardware: string) {
+  return summary.hardwareStatuses.find((status) => status.hardware === hardware);
+}
+
+/**
+ * Leader of the model's first comparable cohort. Leadership only ever exists
+ * inside a cohort, so every ranking assertion reads it from there — there is no
+ * model-global winner to compare across engine families or deployment modes.
+ */
+function primaryLeaderOf(summary: OverviewModelSummary) {
+  return summary.comparisonGroups[0]?.primaryRanking.leader ?? null;
+}
+
+describe('selectOverviewPrecision', () => {
+  it('returns null without any FP4/FP8 rows', () => {
+    expect(selectOverviewPrecision(Model.Qwen3_5, [row({ precision: Precision.BF16 })])).toBeNull();
+    expect(selectOverviewPrecision(Model.Qwen3_5, [row({ precision: Precision.INT4 })])).toBeNull();
+  });
+
+  it('ignores standard-decode rows when counting exact-@50 coverage', () => {
+    // FP8 has the most curves but every one is standard decode; FP4's single
+    // speculative curve is the only exact-@50 coverage, so it wins outnumbered.
+    const rows = [
+      ...frontier([1200, 1000, 800, 600], { hardware: 'b200', precision: Precision.FP4 }),
+      ...frontier([1100, 900, 700, 500], {
+        hardware: 'gb200',
+        precision: Precision.FP8,
+        spec_method: 'none',
+      }),
+      ...frontier([1000, 800, 600, 400], {
+        hardware: 'gb300',
+        precision: Precision.FP8,
+        spec_method: 'none',
+      }),
+    ];
+
+    expect(selectOverviewPrecision(Model.Qwen3_5, rows)).toBe(Precision.FP4);
+  });
+});
+
+describe('buildOverviewModelSummary', () => {
+  it('keeps Kimi point releases as separate exact results, never one blended frontier', () => {
+    // kimik2.5 and kimik2.7-code both map to Model.Kimi_K2_5. Otherwise
+    // identical (same hardware/date/config), a model-blind key would merge them
+    // into one frontier — kimik2.5's 100 tok/s/user point would make the winner
+    // reachable there. Exact identity keeps kimik2.7-code's frontier pure.
+    const summary = buildOverviewModelSummary(Model.Kimi_K2_5, [
+      row({
+        model: 'kimik2.7-code',
+        conc: 1,
+        metrics: { median_intvty: 30, output_tput_per_gpu: 1200 },
+      }),
+      row({
+        model: 'kimik2.7-code',
+        conc: 2,
+        metrics: { median_intvty: 50, output_tput_per_gpu: 1000 },
+      }),
+      row({
+        model: 'kimik2.7-code',
+        conc: 3,
+        metrics: { median_intvty: 75, output_tput_per_gpu: 800 },
+      }),
+      row({ model: 'kimik2.5', conc: 4, metrics: { median_intvty: 50, output_tput_per_gpu: 400 } }),
+      row({
+        model: 'kimik2.5',
+        conc: 5,
+        metrics: { median_intvty: 100, output_tput_per_gpu: 700 },
+      }),
+    ]);
+
+    const leader = primaryLeaderOf(summary);
+
+    expect(leader?.dbModel).toBe('kimik2.7-code');
+    expect(overviewPrimaryValue(leader!)).toBe(1000);
+    expect(leader?.tierValues.find(({ tier }) => tier === 100)).toEqual({
+      tier: 100,
+      value: null,
+      boundary: 'unreachable',
+      evidenceDate: null,
+    });
+  });
+
+  it('splits cohorts by db model so point releases never rank against each other', () => {
+    // Same hardware, engine group and deployment mode; only the db model
+    // differs, so the two point releases must land in separate cohorts.
+    const summary = buildOverviewModelSummary(Model.Kimi_K2_5, [
+      ...frontier([1200, 1000, 800, 600], { model: 'kimik2.5' }),
+      ...frontier([1100, 900, 700, 500], { model: 'kimik2.7-code' }),
+    ]);
+
+    expect(summary.comparisonGroups.map((group) => group.dbModel).toSorted()).toEqual([
+      'kimik2.5',
+      'kimik2.7-code',
+    ]);
+  });
+
+  it('never compares incompatible DeepSeek MTP engine families', () => {
+    const summary = buildOverviewModelSummary(Model.DeepSeek_V4_Pro, [
+      ...frontier([1200, 1000, 800, 600], { framework: 'dynamo-trt' }),
+      ...frontier([1100, 900, 700, 500], { framework: 'mori-sglang' }),
+    ]);
+
+    expect(summary.comparisonGroups.map(({ engineGroup }) => engineGroup).toSorted()).toEqual([
+      'sglang',
+      'trt',
+    ]);
+    expect(summary.comparisonGroups.every((group) => group.primaryRanking.runnerUp === null)).toBe(
+      true,
+    );
+  });
+
+  it('keeps aggregate and disaggregate denominators in separate cohorts', () => {
+    const summary = buildOverviewModelSummary(Model.Qwen3_5, [
+      ...frontier([1200, 1000, 800, 600], { disagg: false }),
+      ...frontier([1800, 1500, 1200, 900], {
+        hardware: 'gb200',
+        disagg: true,
+        is_multinode: true,
+      }),
+    ]);
+    expect(summary.comparisonGroups.map(({ deploymentMode }) => deploymentMode).toSorted()).toEqual(
+      ['aggregated', 'disaggregated'],
+    );
+  });
+
+  it('counts aggregated GPUs as one pool and disaggregated as prefill + decode', () => {
+    const summary = buildOverviewModelSummary(Model.Qwen3_5, [
+      ...frontier([1200, 1000, 800, 600], { num_prefill_gpu: 4, num_decode_gpu: 4 }),
+      ...frontier([1800, 1500, 1200, 900], {
+        hardware: 'gb200',
+        disagg: true,
+        is_multinode: true,
+        num_prefill_gpu: 16,
+        num_decode_gpu: 16,
+      }),
+    ]);
+    const leaderFor = (mode: string) =>
+      summary.comparisonGroups.find((group) => group.deploymentMode === mode)?.primaryRanking
+        .leader;
+
+    expect(leaderFor('aggregated')?.totalGpu).toBe(4);
+    expect(leaderFor('disaggregated')?.totalGpu).toBe(32);
+  });
+
+  it('ranks 50 and 100 independently across hardware with a real gap', () => {
+    // b200 leads 50 (1000 vs 800) but bottoms out at 100 (300), where gb200's
+    // shallower curve (400) takes over — two exact reads, so the gap is real.
+    const summary = buildOverviewModelSummary(Model.Qwen3_5, [
+      ...frontier([1200, 1000, 800, 300], { hardware: 'b200' }),
+      ...frontier([1100, 800, 600, 400], { hardware: 'gb200' }),
+    ]);
+    const [group] = summary.comparisonGroups;
+    const expectedHighConfigKey = overviewConfigIdentityKey(row({ hardware: 'gb200' }));
+
+    expect(group.primaryRanking.state).toBe('comparable');
+    expect(group.primaryRanking.leader?.hardware).toBe('b200');
+    expect(group.primaryRanking.runnerUp?.hardware).toBe('gb200');
+    expect(group.primaryRanking.gapPercent).toBeCloseTo((1000 / 800 - 1) * 100);
+    expect(group.highRanking.leader?.key).toBe(expectedHighConfigKey);
+    expect(group.highRanking.leader?.hardware).toBe('gb200');
+    expect(group.highLeaderTransition).toBe('changed_hardware');
+  });
+
+  it('picks the best per-hardware config independently at each tier', () => {
+    // On b200, sglang wins 50 (1000 > 700) while dynamo-trt wins 100 (400 > 200).
+    const summary = buildOverviewModelSummary(Model.Qwen3_5, [
+      ...frontier([1200, 1000, 800, 200], { framework: 'sglang' }),
+      ...frontier([1100, 700, 500, 400], { framework: 'dynamo-trt' }),
+    ]);
+    const [group] = summary.comparisonGroups;
+
+    expect(group.primaryRanking.leader).toMatchObject({ hardware: 'b200', framework: 'sglang' });
+    expect(group.highRanking.leader).toMatchObject({ hardware: 'b200', framework: 'dynamo-trt' });
+    expect(group.highLeaderTransition).toBe('same_hardware');
+  });
+
+  it('reports no_primary_baseline when 50 has no exact read but 100 does', () => {
+    // Frontier floor at 60 tok/s/user: tier 50 is clamped_low (no exact read),
+    // so the primary tier has no leader, while tier 100 sits inside [60, 110]
+    // and interpolates — a high leader with no primary baseline to compare to.
+    // FP4 rows so this insufficient-coverage precision is the primary (0-0 exact
+    // tie → FP4); the transition mechanics under test are precision-agnostic.
+    const summary = buildOverviewModelSummary(Model.Qwen3_5, [
+      ...frontierAt(
+        [
+          [60, 900],
+          [75, 800],
+          [90, 700],
+          [110, 600],
+        ],
+        { precision: Precision.FP4 },
+      ),
+    ]);
+    const [group] = summary.comparisonGroups;
+
+    expect(group.primaryRanking.state).toBe('insufficient_coverage');
+    expect(group.highLeaderTransition).toBe('no_primary_baseline');
+  });
+
+  it('never lets clamped or unreachable reads lead or form a gap at a tier', () => {
+    const summary = buildOverviewModelSummary(Model.Qwen3_5, [
+      // Frontier floor at 60 tok/s/user → 50 read is clamped_low.
+      ...frontierAt(
+        [
+          [60, 900],
+          [70, 800],
+          [80, 700],
+          [90, 600],
+        ],
+        { hardware: 'b300', precision: Precision.FP4 },
+      ),
+      // Frontier ceiling at 35 tok/s/user → 50 read is unreachable.
+      ...frontierAt(
+        [
+          [20, 500],
+          [25, 450],
+          [30, 400],
+          [35, 350],
+        ],
+        { hardware: 'b200', precision: Precision.FP4 },
+      ),
+    ]);
+    const [group] = summary.comparisonGroups;
+
+    expect(group.primaryRanking.state).toBe('insufficient_coverage');
+    expect(group.primaryRanking.gapPercent).toBeNull();
+    expect(group.primaryRanking.leader).toBeNull();
+    expect(group.hardwareStatuses.find((s) => s.hardware === 'b300')?.primary.boundary).toBe(
+      'clamped_low',
+    );
+  });
+
+  it('lets a lower exact read outrank a higher clamped one', () => {
+    // b300's clamped 900 is its frontier floor, not a 50 tok/s/user result, so
+    // gb200's real 500 leads and no gap is claimed against unmeasured coverage.
+    const summary = buildOverviewModelSummary(Model.Qwen3_5, [
+      ...frontierAt(
+        [
+          [60, 900],
+          [70, 800],
+          [80, 700],
+          [90, 600],
+        ],
+        { hardware: 'b300' },
+      ),
+      ...frontier([700, 500, 400, 300], { hardware: 'gb200' }),
+    ]);
+    const [group] = summary.comparisonGroups;
+
+    expect(group.primaryRanking.leader?.hardware).toBe('gb200');
+    expect(group.primaryRanking.state).toBe('single_measured');
+    expect(group.primaryRanking.gapPercent).toBeNull();
+    expect(group.hardwareStatuses.find((s) => s.hardware === 'b300')?.primary).toMatchObject({
+      value: 900,
+      boundary: 'clamped_low',
+    });
+  });
+
+  it('dates a model by its newest workload row, and null when it has none', () => {
+    const summary = buildOverviewModelSummary(Model.Qwen3_5, [
+      ...frontier([1200, 1000, 800, 600], { date: '2026-07-01' }),
+      ...frontier([1100, 900, 700, 500], { hardware: 'gb200', date: '2026-07-15' }),
+    ]);
+
+    expect(summary.latestWorkloadDate).toBe('2026-07-15');
+    expect(buildOverviewModelSummary(Model.Qwen3_5, []).latestWorkloadDate).toBeNull();
+  });
+});
+
+describe('overview precision selection and secondary', () => {
+  it('makes the wider exact-@50 coverage the primary precision', () => {
+    const summary = buildOverviewModelSummary(Model.Qwen3_5, [
+      ...frontier([1200, 1000, 800, 600], { hardware: 'b200', precision: Precision.FP4 }),
+      ...frontier([1100, 900, 700, 500], { hardware: 'gb200', precision: Precision.FP8 }),
+      ...frontier([1000, 800, 600, 400], { hardware: 'gb300', precision: Precision.FP8 }),
+    ]);
+
+    expect(summary.selectedPrecision).toBe(Precision.FP8);
+  });
+
+  it('breaks an exact-@50 coverage tie toward FP4', () => {
+    const summary = buildOverviewModelSummary(Model.Qwen3_5, [
+      ...frontier([1200, 1000, 800, 600], { hardware: 'b200', precision: Precision.FP4 }),
+      ...frontier([1100, 900, 700, 500], { hardware: 'gb200', precision: Precision.FP8 }),
+    ]);
+
+    expect(summary.selectedPrecision).toBe(Precision.FP4);
+  });
+
+  it('ranks each precision within itself, never across FP4 and FP8', () => {
+    const summary = buildOverviewModelSummary(Model.Qwen3_5, [
+      ...frontier([1200, 1000, 800, 600], { hardware: 'b300', precision: Precision.FP4 }),
+      ...frontier([1100, 900, 700, 500], { hardware: 'b200', precision: Precision.FP4 }),
+      ...frontier([1000, 800, 600, 400], { hardware: 'gb200', precision: Precision.FP8 }),
+    ]);
+
+    // The FP4 cohort ranks only its own two hardware; the FP8 platform is the
+    // secondary precision, never a cross-precision delta in the primary cohort.
+    expect(summary.comparisonGroups[0].hardwareStatuses.map((s) => s.hardware).toSorted()).toEqual([
+      'b200',
+      'b300',
+    ]);
+    expect(summary.secondary?.precision).toBe(Precision.FP8);
+  });
+
+  it('keeps FP8 visible as secondary coverage when FP4 is primary', () => {
+    const summary = buildOverviewModelSummary(Model.Qwen3_5, [
+      ...frontier([1200, 1000, 800, 600], { hardware: 'b300', precision: Precision.FP4 }),
+      ...frontier([1100, 900, 700, 500], { hardware: 'b200', precision: Precision.FP4 }),
+      ...frontier([1000, 800, 600, 400], { hardware: 'h200', precision: Precision.FP8 }),
+    ]);
+
+    expect(summary.selectedPrecision).toBe(Precision.FP4);
+    expect(summary.secondary?.state).toBe('coverage');
+    expect(summary.secondary?.measuredHardware).toContain(getHardwareConfig('h200').label);
+  });
+
+  it('opens secondary rows only when FP8 ranks a pair on hardware FP4 lacks', () => {
+    const ranked = buildOverviewModelSummary(Model.Qwen3_5, [
+      ...frontier([1200, 1000, 800, 600], { hardware: 'b300', precision: Precision.FP4 }),
+      ...frontier([1150, 950, 750, 550], { hardware: 'b200', precision: Precision.FP4 }),
+      ...frontier([1100, 900, 700, 500], { hardware: 'gb300', precision: Precision.FP4 }),
+      ...frontier([1000, 800, 600, 400], { hardware: 'gb200', precision: Precision.FP8 }),
+      ...frontier([980, 780, 580, 380], { hardware: 'mi355x', precision: Precision.FP8 }),
+    ]);
+    // FP4 leads on three hardware; FP8 ranks a comparable pair on two it lacks.
+    expect(ranked.selectedPrecision).toBe(Precision.FP4);
+    expect(ranked.secondary?.state).toBe('ranked');
+
+    const covered = buildOverviewModelSummary(Model.Qwen3_5, [
+      ...frontier([1200, 1000, 800, 600], { hardware: 'gb200', precision: Precision.FP4 }),
+      ...frontier([1150, 950, 750, 550], { hardware: 'mi355x', precision: Precision.FP4 }),
+      ...frontier([900, 850, 650, 450], { hardware: 'gb200', precision: Precision.FP8 }),
+      ...frontier([880, 820, 620, 420], { hardware: 'mi355x', precision: Precision.FP8 }),
+    ]);
+    // Same FP8 pair, but both hardware already rank in FP4 → no new coverage.
+    expect(covered.secondary?.state).toBe('coverage');
+  });
+});
+
+describe('overview not-ranked reasons', () => {
+  it('accounts for every page hardware with an exact value or one reason', () => {
+    const model = [
+      ...frontier([1200, 1000, 800, 600], { hardware: 'b300', precision: Precision.FP4 }),
+      ...frontier([1100, 900, 700, 500], { hardware: 'b200', precision: Precision.FP4 }),
+      // gb200 measured at the primary precision, but standard decode only.
+      ...frontier([1000, 800, 600, 400], {
+        hardware: 'gb200',
+        precision: Precision.FP4,
+        spec_method: 'none',
+      }),
+    ];
+    // mi355x reaches the page order only because another model measured it.
+    const other = frontier([900, 800, 700, 600], { hardware: 'mi355x', precision: Precision.FP4 });
+    const order = buildOverviewHardwareOrder([...model, ...other]);
+    const summary = buildOverviewModelSummary(Model.Qwen3_5, model, order);
+
+    // b300/b200 rank (exact @50) and are absent; gb200 and mi355x each get one.
+    expect(summary.notRanked.map((entry) => [entry.hardware, entry.reason])).toEqual([
+      ['gb200', 'standard_decode_only'],
+      ['mi355x', 'no_8k1k_data'],
+    ]);
+  });
+
+  it('separates a frontier that tops out below 50 from a missing @50 read', () => {
+    const summary = buildOverviewModelSummary(Model.Qwen3_5, [
+      // b200 frontier ceiling at 45 tok/s/user → @50 unreachable.
+      ...frontierAt(
+        [
+          [20, 500],
+          [30, 450],
+          [40, 400],
+          [45, 350],
+        ],
+        { hardware: 'b200', precision: Precision.FP4 },
+      ),
+      // b300 frontier floor at 60 tok/s/user → @50 clamped_low, not unreachable.
+      ...frontierAt(
+        [
+          [60, 900],
+          [70, 800],
+          [80, 700],
+          [90, 600],
+        ],
+        { hardware: 'b300', precision: Precision.FP4 },
+      ),
+    ]);
+    const reasonOf = (hardware: string) =>
+      summary.notRanked.find((entry) => entry.hardware === hardware)?.reason;
+
+    expect(reasonOf('b200')).toBe('cannot_reach_at50');
+    expect(reasonOf('b300')).toBe('no_exact_at50');
+  });
+});
+
+describe('overview result-level evidence dates', () => {
+  it('dates each read from its own config frontier, never a sibling', () => {
+    const summary = buildOverviewModelSummary(Model.Qwen3_5, [
+      ...datedFrontier(
+        [
+          [40, 1100, '2026-07-10'],
+          [60, 900, '2026-07-10'],
+        ],
+        { hardware: 'b300', precision: Precision.FP4 },
+      ),
+      ...datedFrontier(
+        [
+          [40, 900, '2026-07-16'],
+          [60, 700, '2026-07-16'],
+        ],
+        { hardware: 'b200', precision: Precision.FP4 },
+      ),
+    ]);
+
+    expect(statusOf(summary, 'b300')?.primary.evidenceDate).toEqual({
+      from: '2026-07-10',
+      to: '2026-07-10',
+    });
+    expect(statusOf(summary, 'b200')?.primary.evidenceDate).toEqual({
+      from: '2026-07-16',
+      to: '2026-07-16',
+    });
+  });
+
+  it('collapses a same-day bracket and spans a cross-day one', () => {
+    const summary = buildOverviewModelSummary(Model.Qwen3_5, [
+      ...datedFrontier(
+        [
+          [40, 1100, '2026-07-10'],
+          [60, 900, '2026-07-10'],
+        ],
+        { hardware: 'b300', precision: Precision.FP4 },
+      ),
+      ...datedFrontier(
+        [
+          [40, 900, '2026-06-24'],
+          [60, 700, '2026-07-04'],
+        ],
+        { hardware: 'b200', precision: Precision.FP4 },
+      ),
+    ]);
+
+    expect(statusOf(summary, 'b300')?.primary.evidenceDate).toEqual({
+      from: '2026-07-10',
+      to: '2026-07-10',
+    });
+    expect(statusOf(summary, 'b200')?.primary.evidenceDate).toEqual({
+      from: '2026-06-24',
+      to: '2026-07-04',
+    });
+  });
+
+  it('ignores a hidden slower config even when it is newer', () => {
+    const summary = buildOverviewModelSummary(Model.Qwen3_5, [
+      ...datedFrontier(
+        [
+          [40, 1100, '2026-07-10'],
+          [60, 900, '2026-07-10'],
+        ],
+        { hardware: 'b300', precision: Precision.FP4, framework: 'sglang' },
+      ),
+      // Slower same-hardware config, newer runs — never backs the visible read.
+      ...datedFrontier(
+        [
+          [40, 700, '2026-07-20'],
+          [60, 500, '2026-07-20'],
+        ],
+        { hardware: 'b300', precision: Precision.FP4, framework: 'dynamo-trt' },
+      ),
+    ]);
+
+    expect(statusOf(summary, 'b300')?.primary.evidenceDate).toEqual({
+      from: '2026-07-10',
+      to: '2026-07-10',
+    });
+  });
+
+  it('dates @50 and @100 from their own bracketing knots', () => {
+    const summary = buildOverviewModelSummary(Model.Qwen3_5, [
+      ...datedFrontier(
+        [
+          [40, 1200, '2026-07-01'],
+          [60, 1000, '2026-07-01'],
+          [90, 700, '2026-07-15'],
+          [110, 500, '2026-07-15'],
+        ],
+        { hardware: 'b300', precision: Precision.FP4 },
+      ),
+    ]);
+    const status = statusOf(summary, 'b300');
+
+    expect(status?.primary.evidenceDate).toEqual({ from: '2026-07-01', to: '2026-07-01' });
+    expect(status?.high.evidenceDate).toEqual({ from: '2026-07-15', to: '2026-07-15' });
+  });
+});
+
+describe('overview hardware coverage', () => {
+  it('holds hardware order fixed against input order and performance changes', () => {
+    const expected = ['gb300', 'gb200', 'b300', 'b200', 'mi355x'];
+    const build = (throughputs: number[], hardware: string[]) => {
+      const rows = hardware.flatMap((hw, index) =>
+        frontier(
+          [
+            throughputs[index] + 200,
+            throughputs[index],
+            throughputs[index] - 200,
+            throughputs[index] - 400,
+          ],
+          { hardware: hw },
+        ),
+      );
+      return buildOverviewModelSummary(Model.Qwen3_5, rows, buildOverviewHardwareOrder(rows));
+    };
+
+    // Same throughputs, hardware fed in the opposite order: the fastest platform
+    // swaps from mi355x to gb300 while the display order must not move.
+    const mi355xLeads = build([600, 700, 800, 900, 1000], expected);
+    const gb300Leads = build([600, 700, 800, 900, 1000], [...expected].toReversed());
+
+    expect(mi355xLeads.hardwareStatuses.map(({ hardware }) => hardware)).toEqual(expected);
+    expect(gb300Leads.hardwareStatuses.map(({ hardware }) => hardware)).toEqual(expected);
+    expect(primaryLeaderOf(mi355xLeads)).toMatchObject({ hardware: 'mi355x' });
+    expect(primaryLeaderOf(gb300Leads)).toMatchObject({ hardware: 'gb300' });
+    expect(gb300Leads.comparisonGroups[0].hardwareStatuses.map(({ hardware }) => hardware)).toEqual(
+      expected,
+    );
+  });
+
+  it('separates unsupported-precision coverage from hardware this model never ran', () => {
+    // Kimi shape: FP4 standard decode on B200, INT4 only on H200. H200 and
+    // MI355X reach the page order only because another model measured them.
+    const kimi = [
+      ...frontier([1200, 1000, 800, 600], {
+        model: 'kimik2.5',
+        hardware: 'b200',
+        precision: Precision.FP4,
+        spec_method: 'none',
+      }),
+      ...frontier([900, 800, 700, 600], {
+        model: 'kimik2.5',
+        hardware: 'h200',
+        precision: Precision.INT4,
+        spec_method: 'none',
+      }),
+    ];
+    const otherModel = [
+      ...frontier([900, 800, 700, 600], { hardware: 'mi355x', precision: Precision.FP4 }),
+      ...frontier([700, 600, 500, 400], { hardware: 'h200', precision: Precision.FP8 }),
+    ];
+    const order = buildOverviewHardwareOrder([...kimi, ...otherModel]);
+    const summary = buildOverviewModelSummary(Model.Kimi_K2_5, kimi, order);
+
+    expect(order.map(({ hardware }) => hardware)).toEqual(['b200', 'mi355x', 'h200']);
+    expect(summary.hardwareStatuses.map(({ hardware }) => hardware)).toEqual([
+      'b200',
+      'mi355x',
+      'h200',
+    ]);
+    expect(statusOf(summary, 'b200')?.coverage.kind).toBe('standard_only');
+    expect(statusOf(summary, 'h200')?.coverage).toMatchObject({
+      kind: 'unsupported_precision_only',
+      availablePrecisions: [Precision.INT4],
+    });
+    expect(statusOf(summary, 'mi355x')?.coverage).toMatchObject({
+      kind: 'no_workload_data',
+      availablePrecisions: [],
+    });
+  });
+});
+
+// Drift guard for the synthetic overview e2e fixture: if the assembler contract
+// changes, the hand-built rows in overview-rows.json must be regenerated, and
+// this fails loudly instead of the e2e stranding empty data. Every expectation
+// is derived by running this same assembler over the fixture — never eyeballed —
+// so it locks the exact 8.7 states overview.cy.ts renders against.
+describe('assembleOverviewPageData over the overview-rows fixture', () => {
+  it('serves every 8.7 precision, coverage and evidence-date state through the real builder', () => {
+    const page = assembleOverviewPageData(
+      overviewRowsFixture as unknown as Record<string, BenchmarkRow[]>,
+    );
+
+    expect(page.models).toHaveLength(DEFAULT_MODELS.size);
+    expect(page.datasetThroughDate).toBe('2026-07-18');
+    expect(page.hardwareOrder.map((entry) => entry.hardware)).toEqual([
+      'gb300',
+      'gb200',
+      'b300',
+      'b200',
+      'mi355x',
+      'h200',
+      'mi325x',
+      'h100',
+    ]);
+
+    // DeepSeek: FP4 primary ranks B300 over B200; FP8 stays visible as one
+    // coverage line (single measured hardware, not a second table); every other
+    // page hardware carries exactly one not-ranked reason, including BOTH clamp
+    // directions; and the leader's @50 read spans two calendar days — the only
+    // cross-day evidence range in the fixture, so the e2e can exercise the range
+    // label the current live dataset never produces.
+    const deepseek = page.models.find((m) => m.model === Model.DeepSeek_V4_Pro);
+    expect(deepseek?.selectedPrecision).toBe(Precision.FP4);
+    const dsCohort = deepseek?.comparisonGroups[0];
+    expect(dsCohort?.primaryRanking.state).toBe('comparable');
+    expect(dsCohort?.primaryRanking.leader?.hardware).toBe('b300');
+    expect(dsCohort?.highRanking.leader?.hardware).toBe('b200');
+    expect(
+      dsCohort?.hardwareStatuses.find((s) => s.hardware === 'b300')?.primary.evidenceDate,
+    ).toEqual({ from: '2026-06-24', to: '2026-07-04' });
+    expect(deepseek?.secondary?.state).toBe('coverage');
+    expect(deepseek?.secondary?.precision).toBe(Precision.FP8);
+    expect(deepseek?.secondary?.measuredHardware).toEqual(['GB200 NVL72']);
+    expect(
+      Object.fromEntries(
+        (deepseek?.notRanked ?? []).map((entry) => [entry.hardware, entry.reason]),
+      ),
+    ).toMatchObject({
+      gb300: 'cannot_reach_at50',
+      mi355x: 'no_exact_at50',
+      h100: 'standard_decode_only',
+      gb200: 'other_precision_only',
+      h200: 'no_8k1k_data',
+    });
+
+    // MiniMax: the wider exact-@50 FP8 coverage flips the primary precision to
+    // FP8 (the coverage rule), dropping FP4 to a single-hardware coverage line.
+    const minimax = page.models.find((m) => m.model === Model.MiniMax_M3);
+    expect(minimax?.selectedPrecision).toBe(Precision.FP8);
+    expect(minimax?.comparisonGroups[0]?.primaryRanking.leader?.hardware).toBe('h200');
+    expect(minimax?.secondary?.state).toBe('coverage');
+    expect(minimax?.secondary?.precision).toBe(Precision.FP4);
+    expect(minimax?.secondary?.measuredHardware).toEqual(['H200']);
+
+    // Qwen: FP4 primary, and FP8 secondary earns a FULL ranked block — it ranks a
+    // comparable pair AND adds MI355X, hardware FP4 has no exact @50 read for.
+    const qwen = page.models.find((m) => m.model === Model.Qwen3_5);
+    expect(qwen?.selectedPrecision).toBe(Precision.FP4);
+    expect(qwen?.secondary?.state).toBe('ranked');
+    expect(qwen?.secondary?.precision).toBe(Precision.FP8);
+    const qwenSecondaryCohort = qwen?.secondary?.comparisonGroups[0];
+    expect(qwenSecondaryCohort?.primaryRanking.state).toBe('comparable');
+    expect(qwenSecondaryCohort?.primaryRanking.leader?.hardware).toBe('b200');
+    expect(qwen?.secondary?.measuredHardware).toEqual(['B200', 'MI355X']);
+  });
+});
diff --git a/packages/app/src/lib/overview-data.ts b/packages/app/src/lib/overview-data.ts
new file mode 100644
index 00000000..6a4e4d59
--- /dev/null
+++ b/packages/app/src/lib/overview-data.ts
@@ -0,0 +1,969 @@
+import { resolveFrameworkPartLabel } from '@semianalysisai/inferencex-constants';
+
+import { parallelismLabel } from '@/components/inference/utils/parallelism-label';
+
+import type { BenchmarkRow } from './api';
+import { buildAvailabilityHwKey } from './chart-utils';
+import { getHardwareConfig, getModelSortIndex } from './constants';
+import {
+  DEFAULT_MODELS,
+  getModelExclusion,
+  getModelLabel,
+  Precision,
+  type Model,
+} from './data-mappings';
+import { buildExclusion, type Exclusion } from './exclusion';
+import { overviewConfigIdentityKey } from './overview-config-identity';
+import { computeTcoFeed, type TcoTierBoundary } from './tco-feed';
+
+export const OVERVIEW_WORKLOAD = { isl: 8192, osl: 1024 } as const;
+export const OVERVIEW_TIERS = [30, 50, 75, 100] as const;
+/** Headline service point every platform is ranked at. */
+export const OVERVIEW_PRIMARY_TIER = 50;
+/** High-interactivity capability read; also drives the high-tier leader transition. */
+export const OVERVIEW_HIGH_TIER = 100;
+
+export interface OverviewTierValue {
+  tier: number;
+  value: number | null;
+  boundary: TcoTierBoundary;
+  /**
+   * Run dates of the frontier point(s) backing this value — the two bracketing
+   * points when interpolated (from = earlier, to = later; equal on the same
+   * day), the single point twice when clamped, null when `value` is null. Comes
+   * only from this config's own frontier at this tier, never a sibling or cohort.
+   */
+  evidenceDate: { from: string; to: string } | null;
+}
+
+/**
+ * One real deployable serving configuration, identified by its exact
+ * deployment topology (model × hardware × framework × precision × spec_method ×
+ * disagg × multinode × per-role parallelism × GPU counts × offload). Tier
+ * values come from this configuration's own Pareto frontier only — never
+ * blended across configurations.
+ */
+export interface OverviewConfigResult {
+  key: string;
+  dbModel: string;
+  hardware: string;
+  hardwareLabel: string;
+  hwKey: string;
+  framework: string;
+  frameworkLabel: string;
+  specMethod: string;
+  specLabel: string;
+  disagg: boolean;
+  precision: string;
+  offloadMode: string;
+  isMultinode: boolean;
+  numPrefillGpu: number;
+  numDecodeGpu: number;
+  /** Physical GPUs backing this config: one shared pool when aggregated, prefill
+   * plus decode when disaggregated. */
+  totalGpu: number;
+  parallelism: string;
+  image: string | null;
+  sourceRunUrls: string[];
+  tierValues: OverviewTierValue[];
+  latestDate: string;
+  oldestFrontierDate: string;
+}
+
+/** One hardware's frontier read at a single tier, with its backing config. */
+export interface OverviewTierRead {
+  tier: number;
+  value: number | null;
+  boundary: TcoTierBoundary | null;
+  /** Copied from the backing config's tier value; null when there is no read. */
+  evidenceDate: { from: string; to: string } | null;
+  config: OverviewConfigResult | null;
+}
+
+/**
+ * Why a hardware is or is not competing in the model's ranking. Coverage is
+ * derived from the raw workload rows alone, so a hardware never disappears just
+ * because a filter or a ranking excluded it:
+ *
+ *  - `comparable_spec` — speculative results at the selected precision, ranked;
+ *  - `standard_only` — selected precision measured, but standard decode only;
+ *  - `alternate_precision_only` — measured at the other FP4/FP8 precision;
+ *  - `other_engine_group_only` — speculative, but only in an engine
+ *    comparability group no other hardware here shares;
+ *  - `unsupported_precision_only` — measured only outside FP4/FP8;
+ *  - `no_workload_data` — no 8K/1K single-turn row for this model.
+ */
+export type OverviewCoverageKind =
+  | 'comparable_spec'
+  | 'standard_only'
+  | 'alternate_precision_only'
+  | 'other_engine_group_only'
+  | 'unsupported_precision_only'
+  | 'no_workload_data';
+
+export interface OverviewHardwareCoverage {
+  hardware: string;
+  hardwareLabel: string;
+  kind: OverviewCoverageKind;
+  availablePrecisions: string[];
+}
+
+/** One hardware in the page-wide display order, which never encodes performance. */
+export interface OverviewHardwareOrderEntry {
+  hardware: string;
+  hardwareLabel: string;
+}
+
+export interface OverviewHardwareStatus {
+  hardware: string;
+  hardwareLabel: string;
+  coverage: OverviewHardwareCoverage;
+  /**
+   * Primary-tier read of this hardware's best configuration in the surrounding
+   * config set — the same read the ranking used, so a row can never display a
+   * number the ranking did not. Null read when the hardware has no ranked
+   * configuration; `coverage.kind` says why.
+   */
+  primary: OverviewTierRead;
+  /** High-tier read, taken from this hardware's own best config at that tier. */
+  high: OverviewTierRead;
+  isPrimaryLeader: boolean;
+  /** Signed percentage against the primary leader; null for the leader itself
+   * and whenever either side is not an exact in-range read. */
+  primaryDeltaPercent: number | null;
+}
+
+/** Whether a cohort's throughput denominator is aggregated or disaggregated. */
+export type OverviewDeploymentMode = 'aggregated' | 'disaggregated';
+
+/**
+ * Ranking state at one tier: `comparable` (≥2 exact hardware reads, gap
+ * computable), `single_measured` (one exact read), or `insufficient_coverage`
+ * (no exact read — every read is clamped or unreachable).
+ */
+export type OverviewRankingState = 'comparable' | 'single_measured' | 'insufficient_coverage';
+
+export interface OverviewTierRanking {
+  tier: number;
+  state: OverviewRankingState;
+  leader: OverviewConfigResult | null;
+  runnerUp: OverviewConfigResult | null;
+  gapPercent: number | null;
+}
+
+/**
+ * One comparable cohort: exact configs that share an engine comparability group
+ * (per the model's exclusion policy) and a deployment mode. Tiers are ranked
+ * independently inside a cohort; configs never compare across cohorts.
+ */
+export interface OverviewComparisonGroup {
+  id: string;
+  dbModel: string;
+  engineGroup: string | null;
+  deploymentMode: OverviewDeploymentMode;
+  hardwareStatuses: OverviewHardwareStatus[];
+  primaryRanking: OverviewTierRanking;
+  highRanking: OverviewTierRanking;
+  /**
+   * How the high-tier leader relates to the primary-tier leader: a new hardware
+   * took the lead (`changed_hardware`), the same hardware held it
+   * (`same_hardware`), there was no primary leader to compare against
+   * (`no_primary_baseline`), or the high tier has no leader at all (`null`).
+   */
+  highLeaderTransition: 'same_hardware' | 'changed_hardware' | 'no_primary_baseline' | null;
+}
+
+export interface OverviewSecondaryPrecision {
+  precision: string;
+  /** 'ranked' → full rows render; 'coverage' → one compact line renders. */
+  state: 'ranked' | 'coverage';
+  /** Populated when state === 'ranked'; [] otherwise. Built by the SAME pipeline as the primary (cohorts, rankings, transitions) over the secondary precision's rows. */
+  comparisonGroups: OverviewComparisonGroup[];
+  hardwareStatuses: OverviewHardwareStatus[];
+  /** Hardware labels measured at this precision (any decode mode), page order. Always populated. */
+  measuredHardware: string[];
+}
+
+/**
+ * Why a page hardware carries no visible value for a model. Derived from the
+ * existing coverage kinds and ranking outcomes, never a parallel truth:
+ *
+ *  - `standard_decode_only` — from `standard_only`;
+ *  - `other_precision_only` — from `alternate_precision_only` and unranked by
+ *    the secondary block; the renderer prints 'FP8 only'/'FP4 only';
+ *  - `int4_bf16_only` — from `unsupported_precision_only`;
+ *  - `different_serving_cohort` — from `other_engine_group_only`;
+ *  - `no_8k1k_data` — from `no_workload_data`;
+ *  - `cannot_reach_at50` — comparable spec whose frontier tops out below 50;
+ *  - `no_exact_at50` — comparable spec with no exact @50 read for any other reason.
+ */
+export type OverviewNotRankedReason =
+  | 'standard_decode_only'
+  | 'other_precision_only'
+  | 'int4_bf16_only'
+  | 'different_serving_cohort'
+  | 'no_8k1k_data'
+  | 'cannot_reach_at50'
+  | 'no_exact_at50';
+
+export interface OverviewNotRankedEntry {
+  hardware: string;
+  hardwareLabel: string;
+  reason: OverviewNotRankedReason;
+  /** FP4/FP8 precisions this hardware measured, for the 'FP8 only' style copy. */
+  precisions: string[];
+}
+
+export interface OverviewModelSummary {
+  model: Model;
+  modelLabel: string;
+  /**
+   * Primary precision: whichever of FP4/FP8 covers more unique hardware with an
+   * exact @50 speculative read; a tie (including 0-0 when FP4/FP8 rows exist)
+   * goes to FP4. Null only when the model has no FP4/FP8 workload rows at all.
+   */
+  selectedPrecision: string | null;
+  /**
+   * Every page hardware's coverage and tier reads for this model, in the
+   * page-wide order. Leadership is claimed here only when a single cohort
+   * exists; with several cohorts the claim belongs to `comparisonGroups`.
+   */
+  hardwareStatuses: OverviewHardwareStatus[];
+  /** Comparable cohorts, each ranked independently at the primary and high tiers. */
+  comparisonGroups: OverviewComparisonGroup[];
+  /** The other of FP4/FP8, or null when it has no workload rows. Ranks only
+   * within itself — never against the primary precision. */
+  secondary: OverviewSecondaryPrecision | null;
+  /**
+   * One entry per page hardware with no visible value — neither an exact @50
+   * read in the primary block nor a ranked read in a `ranked` secondary. Every
+   * page hardware is therefore accounted for per model: an exact value, or one
+   * not-ranked reason.
+   */
+  notRanked: OverviewNotRankedEntry[];
+  /**
+   * Newest workload row measured for this model. Now that every visible result
+   * carries its own `evidenceDate`, this exists ONLY to date a coverage-only
+   * model with zero visible results. Null when the model has no workload row.
+   */
+  latestWorkloadDate: string | null;
+  emptyReason: 'no_8k1k_single_turn_data' | 'no_fp4_fp8_data' | null;
+}
+
+export interface OverviewPageData {
+  models: OverviewModelSummary[];
+  hardwareOrder: { hardware: string; hardwareLabel: string }[];
+  datasetThroughDate: string | null;
+}
+
+/** Precisions the overview may rank, in preference order. */
+const OVERVIEW_PRECISIONS: readonly string[] = [Precision.FP4, Precision.FP8];
+
+/** Rows measuring the fixed overview workload, before any other filter. */
+function overviewWorkloadRows(rows: readonly BenchmarkRow[]): BenchmarkRow[] {
+  return rows.filter(
+    (row) =>
+      row.benchmark_type === 'single_turn' &&
+      row.isl === OVERVIEW_WORKLOAD.isl &&
+      row.osl === OVERVIEW_WORKLOAD.osl,
+  );
+}
+
+/**
+ * Page-wide hardware display order: every base hardware with FP4/FP8 coverage
+ * at the overview workload, in hardware-registry order. Built before precision
+ * selection and before any ranking, so neither a precision fallback nor a
+ * measurement change can reorder or drop a column.
+ */
+export function buildOverviewHardwareOrder(
+  rows: readonly BenchmarkRow[],
+): OverviewHardwareOrderEntry[] {
+  const hardware = new Set(
+    overviewWorkloadRows(rows)
+      .filter((row) => OVERVIEW_PRECISIONS.includes(row.precision))
+      .map((row) => row.hardware),
+  );
+  return [...hardware]
+    .toSorted((a, b) => getModelSortIndex(a) - getModelSortIndex(b) || a.localeCompare(b))
+    .map((entry) => ({ hardware: entry, hardwareLabel: getHardwareConfig(entry).label }));
+}
+
+/**
+ * Newest raw workload row across the page. Deliberately not derived from the
+ * retained winners: a precision or engine the page never ranks still dates the
+ * dataset it was measured in.
+ */
+export function overviewDatasetThroughDate(rows: readonly BenchmarkRow[]): string | null {
+  return overviewWorkloadRows(rows).reduce<string | null>(
+    (latest, row) => (latest === null || row.date > latest ? row.date : latest),
+    null,
+  );
+}
+
+/** Newest date among a model's own workload rows; null when it has none. */
+function latestWorkloadDateOf(workloadRows: readonly BenchmarkRow[]): string | null {
+  return workloadRows.reduce<string | null>(
+    (latest, row) => (latest === null || row.date > latest ? row.date : latest),
+    null,
+  );
+}
+
+/**
+ * Speculative configs for one precision, grouped by exact deployment identity.
+ * Standard-decode rows never enter the pool, so a precision is ranked on its
+ * speculative frontier alone.
+ */
+function buildPrecisionConfigs(
+  model: Model,
+  workloadRows: readonly BenchmarkRow[],
+  precision: string,
+): OverviewConfigResult[] {
+  const rowsByConfig = new Map<string, BenchmarkRow[]>();
+  for (const row of workloadRows) {
+    if (row.precision !== precision || row.spec_method === 'none') continue;
+    const key = overviewConfigIdentityKey(row);
+    const configRows = rowsByConfig.get(key);
+    if (configRows) configRows.push(row);
+    else rowsByConfig.set(key, [row]);
+  }
+
+  const configs: OverviewConfigResult[] = [];
+  for (const [key, configRows] of rowsByConfig) {
+    const config = buildConfigResult(model, precision, key, configRows);
+    if (config) configs.push(config);
+  }
+  return configs;
+}
+
+/** Unique hardware whose best @50 read is exact (in-range) among `configs`. */
+function exactPrimaryHardware(configs: readonly OverviewConfigResult[]): Set<string> {
+  return new Set(
+    [...readsByHardwareAtTier(configs, OVERVIEW_PRIMARY_TIER).values()]
+      .filter(isExactTierRead)
+      .map((read) => read.config.hardware),
+  );
+}
+
+/** Whichever precision covers more exact-@50 hardware; a tie (incl. 0-0) → FP4. */
+function selectPrimaryFrom(
+  configsByPrecision: ReadonlyMap<string, readonly OverviewConfigResult[]>,
+): string {
+  const exactCount = (precision: string) =>
+    exactPrimaryHardware(configsByPrecision.get(precision) ?? []).size;
+  return exactCount(Precision.FP8) > exactCount(Precision.FP4) ? Precision.FP8 : Precision.FP4;
+}
+
+/**
+ * Primary precision by exact-@50 hardware coverage: FP4 and FP8 are built and
+ * counted independently, the wider coverage wins, and a tie (including 0-0 when
+ * FP4/FP8 rows exist) goes to FP4. Null only when no FP4/FP8 workload row exists
+ * — coverage volume of one vendor's curves can shift which precision is primary,
+ * but never which precisions are ranked.
+ */
+export function selectOverviewPrecision(
+  model: Model,
+  workloadRows: readonly BenchmarkRow[],
+): string | null {
+  if (!workloadRows.some((row) => OVERVIEW_PRECISIONS.includes(row.precision))) return null;
+  const configsByPrecision = new Map(
+    OVERVIEW_PRECISIONS.map(
+      (precision) => [precision, buildPrecisionConfigs(model, workloadRows, precision)] as const,
+    ),
+  );
+  return selectPrimaryFrom(configsByPrecision);
+}
+
+function tierValueAt(config: OverviewConfigResult, tier: number): number | null {
+  return config.tierValues.find((value) => value.tier === tier)?.value ?? null;
+}
+
+export function overviewPrimaryValue(config: OverviewConfigResult): number | null {
+  return tierValueAt(config, OVERVIEW_PRIMARY_TIER);
+}
+
+export function overviewHighValue(config: OverviewConfigResult): number | null {
+  return tierValueAt(config, OVERVIEW_HIGH_TIER);
+}
+
+function readConfigAtTier(config: OverviewConfigResult, tier: number): OverviewTierRead {
+  const tierValue = config.tierValues.find((value) => value.tier === tier);
+  return {
+    tier,
+    value: tierValue?.value ?? null,
+    boundary: tierValue?.boundary ?? null,
+    evidenceDate: tierValue?.evidenceDate ?? null,
+    config,
+  };
+}
+
+/** A tier read known to be backed by a configuration. */
+interface ConfigTierRead extends OverviewTierRead {
+  config: OverviewConfigResult;
+}
+
+/**
+ * A tier read counts only when it lands inside the configuration's own measured
+ * frontier. A clamped or unreachable read is a coverage gap, so it can never
+ * lead a tier or anchor a percentage gap.
+ */
+const isExactTierRead = <T extends OverviewTierRead>(read: T): read is T & { value: number } =>
+  read.value !== null && read.boundary === 'interpolated';
+
+function compareTierReads(a: ConfigTierRead, b: ConfigTierRead): number {
+  return (
+    (b.value ?? -1) - (a.value ?? -1) ||
+    getModelSortIndex(a.config.hardware) - getModelSortIndex(b.config.hardware) ||
+    a.config.key.localeCompare(b.config.key)
+  );
+}
+
+/**
+ * One read per hardware at `tier`: its best exact in-range read, or — only when
+ * the hardware has no exact read at all — its best out-of-range read, kept so
+ * the hardware still surfaces as a coverage gap. Ranking and hardware rows both
+ * consume this, so a row can never display a number the ranking did not use.
+ */
+function readsByHardwareAtTier(
+  configs: readonly OverviewConfigResult[],
+  tier: number,
+): Map<string, ConfigTierRead> {
+  const reads = configs
+    .map((config): ConfigTierRead => ({ ...readConfigAtTier(config, tier), config }))
+    .toSorted(compareTierReads);
+
+  const byHardware = new Map<string, ConfigTierRead>();
+  for (const read of reads.filter(isExactTierRead)) {
+    if (!byHardware.has(read.config.hardware)) byHardware.set(read.config.hardware, read);
+  }
+  for (const read of reads) {
+    if (!byHardware.has(read.config.hardware)) byHardware.set(read.config.hardware, read);
+  }
+  return byHardware;
+}
+
+function nullTierRead(tier: number): OverviewTierRead {
+  return { tier, value: null, boundary: null, evidenceDate: null, config: null };
+}
+
+/**
+ * Classify one hardware's coverage from its raw workload rows. `comparable`
+ * lists the hardware sharing an engine comparability group with at least one
+ * other hardware; a platform whose only speculative results sit in an engine
+ * family nobody else here uses is reported as coverage rather than a competitor.
+ */
+function coverageKindFor(
+  hardwareRows: readonly BenchmarkRow[],
+  selectedPrecision: string | null,
+  hardware: string,
+  comparable: ReadonlySet<string>,
+): OverviewCoverageKind {
+  if (hardwareRows.length === 0) return 'no_workload_data';
+  const candidates = hardwareRows.filter((row) => OVERVIEW_PRECISIONS.includes(row.precision));
+  if (candidates.length === 0) return 'unsupported_precision_only';
+  const selected = candidates.filter((row) => row.precision === selectedPrecision);
+  if (selected.length === 0) return 'alternate_precision_only';
+  if (!selected.some((row) => row.spec_method !== 'none')) return 'standard_only';
+  return comparable.size === 0 || comparable.has(hardware)
+    ? 'comparable_spec'
+    : 'other_engine_group_only';
+}
+
+/**
+ * One status row per entry of `coverage`, which already carries the stable
+ * hardware order. `configs` scopes the tier reads and the leader: the whole
+ * model's ranked set at model level, one cohort's set inside a comparison group.
+ */
+function buildHardwareStatuses(
+  coverage: readonly OverviewHardwareCoverage[],
+  configs: readonly OverviewConfigResult[],
+  withLeader: boolean,
+): OverviewHardwareStatus[] {
+  const primaryReads = readsByHardwareAtTier(configs, OVERVIEW_PRIMARY_TIER);
+  const highReads = readsByHardwareAtTier(configs, OVERVIEW_HIGH_TIER);
+  const leader = withLeader
+    ? [...primaryReads.values()].filter(isExactTierRead).toSorted(compareTierReads)[0]
+    : undefined;
+
+  return coverage.map((hardwareCoverage) => {
+    const { hardware, hardwareLabel } = hardwareCoverage;
+    const primary = primaryReads.get(hardware) ?? nullTierRead(OVERVIEW_PRIMARY_TIER);
+    const isPrimaryLeader = leader !== undefined && primary.config?.key === leader.config.key;
+    return {
+      hardware,
+      hardwareLabel,
+      coverage: hardwareCoverage,
+      primary,
+      high: highReads.get(hardware) ?? nullTierRead(OVERVIEW_HIGH_TIER),
+      isPrimaryLeader,
+      primaryDeltaPercent:
+        leader !== undefined && !isPrimaryLeader && isExactTierRead(primary) && leader.value > 0
+          ? (primary.value / leader.value - 1) * 100
+          : null,
+    };
+  });
+}
+
+/**
+ * Rank one comparable cohort at a single tier. Every configuration is read at
+ * `tier`, each hardware contributes only its best exact read, and the surviving
+ * hardware are ordered by that read — so each tier is ranked on its own
+ * evidence and a 50 winner is never carried into the 100 ranking.
+ */
+function rankAtTier(configs: readonly OverviewConfigResult[], tier: number): OverviewTierRanking {
+  const hardwareLeaders = [...readsByHardwareAtTier(configs, tier).values()]
+    .filter(isExactTierRead)
+    .toSorted(compareTierReads);
+
+  const [leader, runnerUp] = hardwareLeaders;
+  return {
+    tier,
+    state:
+      hardwareLeaders.length >= 2
+        ? 'comparable'
+        : hardwareLeaders.length === 1
+          ? 'single_measured'
+          : 'insufficient_coverage',
+    leader: leader?.config ?? null,
+    runnerUp: runnerUp?.config ?? null,
+    gapPercent:
+      leader && runnerUp && runnerUp.value > 0 ? (leader.value / runnerUp.value - 1) * 100 : null,
+  };
+}
+
+interface OverviewCohort {
+  dbModel: string;
+  engineGroup: string | null;
+  deploymentMode: OverviewDeploymentMode;
+  configs: OverviewConfigResult[];
+}
+
+/**
+ * How a cohort's high-tier leader relates to its primary-tier leader. Compared
+ * on hardware, not config identity: a same-hardware engine or precision swap
+ * between tiers is not a hardware change, and a high tier with no primary
+ * baseline to compare against says so rather than implying a change.
+ */
+function computeHighLeaderTransition(
+  primaryRanking: OverviewTierRanking,
+  highRanking: OverviewTierRanking,
+): OverviewComparisonGroup['highLeaderTransition'] {
+  if (highRanking.leader === null) return null;
+  if (primaryRanking.leader === null) return 'no_primary_baseline';
+  return highRanking.leader.hardware === primaryRanking.leader.hardware
+    ? 'same_hardware'
+    : 'changed_hardware';
+}
+
+function buildComparisonGroup(
+  id: string,
+  cohort: OverviewCohort,
+  coverage: readonly OverviewHardwareCoverage[],
+): OverviewComparisonGroup {
+  const cohortHardware = new Set(cohort.configs.map(({ hardware }) => hardware));
+  const primaryRanking = rankAtTier(cohort.configs, OVERVIEW_PRIMARY_TIER);
+  const highRanking = rankAtTier(cohort.configs, OVERVIEW_HIGH_TIER);
+  return {
+    id,
+    dbModel: cohort.dbModel,
+    engineGroup: cohort.engineGroup,
+    deploymentMode: cohort.deploymentMode,
+    hardwareStatuses: buildHardwareStatuses(
+      coverage.filter(({ hardware }) => cohortHardware.has(hardware)),
+      cohort.configs,
+      true,
+    ),
+    primaryRanking,
+    highRanking,
+    highLeaderTransition: computeHighLeaderTransition(primaryRanking, highRanking),
+  };
+}
+
+/**
+ * Partition configurations into cohorts that may actually be compared: only
+ * configs sharing a db model (one display model spans several point releases,
+ * which must never rank against each other), an engine comparability group (the
+ * model's exclusion policy — DeepSeek MTP acceptance forcing differs per engine
+ * family) and a throughput denominator (aggregated vs disaggregated) are ranked
+ * against each other.
+ */
+function buildCohorts(
+  exclusion: Exclusion,
+  configs: readonly OverviewConfigResult[],
+): Map<string, OverviewCohort> {
+  const cohorts = new Map<string, OverviewCohort>();
+  for (const config of configs) {
+    const engineGroup = exclusion.groupOf(config.hwKey);
+    const deploymentMode: OverviewDeploymentMode = config.disagg ? 'disaggregated' : 'aggregated';
+    const id = `${config.dbModel}|${engineGroup ?? 'any'}|${deploymentMode}`;
+    const cohort = cohorts.get(id);
+    if (cohort) cohort.configs.push(config);
+    else
+      cohorts.set(id, { dbModel: config.dbModel, engineGroup, deploymentMode, configs: [config] });
+  }
+  return cohorts;
+}
+
+/**
+ * Hardware sharing an engine comparability group with at least one other
+ * platform. Partitioned by engine group alone, never by the ranking cohorts:
+ * cohorts also split on deployment mode, so a platform measured only
+ * disaggregated would look isolated while running the very engine family its
+ * peers run — and `other_engine_group_only` would then assert a reason that is
+ * false.
+ */
+function comparableHardware(
+  exclusion: Exclusion,
+  configs: readonly OverviewConfigResult[],
+): Set<string> {
+  const byEngineGroup = new Map<string | null, Set<string>>();
+  for (const config of configs) {
+    const group = exclusion.groupOf(config.hwKey);
+    const groupHardware = byEngineGroup.get(group);
+    if (groupHardware) groupHardware.add(config.hardware);
+    else byEngineGroup.set(group, new Set([config.hardware]));
+  }
+  return new Set(
+    [...byEngineGroup.values()]
+      .filter((groupHardware) => groupHardware.size > 1)
+      .flatMap((groupHardware) => [...groupHardware]),
+  );
+}
+
+function buildComparisonGroups(
+  cohorts: ReadonlyMap<string, OverviewCohort>,
+  coverage: readonly OverviewHardwareCoverage[],
+): OverviewComparisonGroup[] {
+  const leadValue = (group: OverviewComparisonGroup): number =>
+    group.primaryRanking.leader ? (overviewPrimaryValue(group.primaryRanking.leader) ?? -1) : -1;
+  return [...cohorts.entries()]
+    .map(([id, cohort]) => buildComparisonGroup(id, cohort, coverage))
+    .toSorted((a, b) => leadValue(b) - leadValue(a) || a.id.localeCompare(b.id));
+}
+
+function buildConfigResult(
+  model: Model,
+  precision: string,
+  key: string,
+  rows: BenchmarkRow[],
+): OverviewConfigResult | null {
+  const feed = computeTcoFeed(rows, [OVERVIEW_WORKLOAD], OVERVIEW_TIERS);
+  if (feed.length === 0) return null;
+
+  const first = rows[0];
+  const { hardware, framework, spec_method: specMethod, disagg } = first;
+  const latestRow = rows.reduce((a, b) => (b.date > a.date ? b : a));
+  const sourceRunUrls = [
+    ...new Set(rows.flatMap((row) => (row.run_url === null ? [] : [row.run_url]))),
+  ].toSorted();
+  return {
+    key,
+    dbModel: first.model,
+    hardware,
+    hardwareLabel: getHardwareConfig(hardware, model).label,
+    hwKey: buildAvailabilityHwKey(hardware, framework, specMethod, disagg),
+    framework,
+    frameworkLabel: resolveFrameworkPartLabel(model, framework),
+    specMethod,
+    specLabel: resolveFrameworkPartLabel(model, specMethod),
+    disagg,
+    precision,
+    offloadMode: first.offload_mode ?? 'off',
+    isMultinode: first.is_multinode,
+    numPrefillGpu: first.num_prefill_gpu,
+    numDecodeGpu: first.num_decode_gpu,
+    totalGpu: disagg ? first.num_prefill_gpu + first.num_decode_gpu : first.num_decode_gpu,
+    parallelism: parallelismLabel({
+      tp: first.decode_tp,
+      ep: first.decode_ep,
+      dpAttention: first.decode_dp_attention,
+      disagg: first.disagg,
+      isMultinode: first.is_multinode,
+      prefillTp: first.prefill_tp,
+      prefillEp: first.prefill_ep,
+      prefillDpAttention: first.prefill_dp_attention,
+      prefillNumWorkers: first.prefill_num_workers,
+      decodeTp: first.decode_tp,
+      decodeEp: first.decode_ep,
+      decodeDpAttention: first.decode_dp_attention,
+      decodeNumWorkers: first.decode_num_workers,
+    }),
+    image: latestRow.image,
+    sourceRunUrls,
+    tierValues: feed.map((row) => {
+      const value =
+        row.boundary === 'unreachable' && row.output_tput_per_gpu === 0
+          ? null
+          : row.output_tput_per_gpu;
+      return {
+        tier: row.tier,
+        value,
+        boundary: row.boundary,
+        evidenceDate: value === null ? null : row.evidence_date,
+      };
+    }),
+    latestDate: feed[0].latest_date,
+    oldestFrontierDate: feed[0].oldest_frontier_date,
+  };
+}
+
+/**
+ * Coverage for every hardware in `order`, derived from raw workload rows only.
+ * Hardware the model never measured stays in the list as `no_workload_data`,
+ * so a platform is never silently absent from one model's row.
+ */
+function buildCoverage(
+  order: readonly OverviewHardwareOrderEntry[],
+  workloadRows: readonly BenchmarkRow[],
+  selectedPrecision: string | null,
+  comparable: ReadonlySet<string>,
+): OverviewHardwareCoverage[] {
+  const rowsByHardware = new Map<string, BenchmarkRow[]>();
+  for (const row of workloadRows) {
+    const hardwareRows = rowsByHardware.get(row.hardware);
+    if (hardwareRows) hardwareRows.push(row);
+    else rowsByHardware.set(row.hardware, [row]);
+  }
+
+  return order.map(({ hardware, hardwareLabel }) => {
+    const hardwareRows = rowsByHardware.get(hardware) ?? [];
+    return {
+      hardware,
+      hardwareLabel,
+      kind: coverageKindFor(hardwareRows, selectedPrecision, hardware, comparable),
+      availablePrecisions: [...new Set(hardwareRows.map((row) => row.precision))].toSorted(),
+    };
+  });
+}
+
+interface OverviewPrecisionBlock {
+  comparisonGroups: OverviewComparisonGroup[];
+  hardwareStatuses: OverviewHardwareStatus[];
+}
+
+/**
+ * Rank one precision's speculative configs: cohorts, per-tier rankings and the
+ * page-wide hardware statuses. The primary and secondary precisions are both
+ * built through here, so neither can rank against the other.
+ */
+function buildPrecisionBlock(
+  model: Model,
+  precision: string,
+  configs: readonly OverviewConfigResult[],
+  hardwareOrder: readonly OverviewHardwareOrderEntry[],
+  workloadRows: readonly BenchmarkRow[],
+): OverviewPrecisionBlock {
+  const exclusion = buildExclusion(getModelExclusion(model));
+  const cohorts = buildCohorts(exclusion, configs);
+  const comparable = comparableHardware(exclusion, configs);
+  const coverage = buildCoverage(hardwareOrder, workloadRows, precision, comparable);
+  const comparisonGroups = buildComparisonGroups(cohorts, coverage);
+  // Leadership is a claim against a comparable cohort. Across several cohorts
+  // there is no model-global leader, so the model row asserts none.
+  return {
+    comparisonGroups,
+    hardwareStatuses: buildHardwareStatuses(coverage, configs, comparisonGroups.length <= 1),
+  };
+}
+
+/**
+ * The other of FP4/FP8. `ranked` (full rows) iff it both ranks a comparable
+ * cohort AND covers hardware the primary precision lacks; otherwise `coverage`
+ * (one compact line, ranked rows suppressed). `measuredHardware` always lists
+ * every hardware measured at this precision so it never disappears from the page.
+ */
+function buildSecondaryPrecision(
+  model: Model,
+  precision: string,
+  configs: readonly OverviewConfigResult[],
+  primaryExact: ReadonlySet<string>,
+  hardwareOrder: readonly OverviewHardwareOrderEntry[],
+  workloadRows: readonly BenchmarkRow[],
+): OverviewSecondaryPrecision | null {
+  const measuredSet = new Set(
+    workloadRows.filter((row) => row.precision === precision).map((row) => row.hardware),
+  );
+  if (measuredSet.size === 0) return null;
+  const measuredHardware = hardwareOrder
+    .filter(({ hardware }) => measuredSet.has(hardware))
+    .map(({ hardwareLabel }) => hardwareLabel);
+
+  const block = buildPrecisionBlock(model, precision, configs, hardwareOrder, workloadRows);
+  const addsCoverage = [...exactPrimaryHardware(configs)].some((hw) => !primaryExact.has(hw));
+  const hasComparable = block.comparisonGroups.some(
+    (group) => group.primaryRanking.state === 'comparable',
+  );
+  const ranked = hasComparable && addsCoverage;
+  return {
+    precision,
+    state: ranked ? 'ranked' : 'coverage',
+    comparisonGroups: ranked ? block.comparisonGroups : [],
+    hardwareStatuses: ranked ? block.hardwareStatuses : [],
+    measuredHardware,
+  };
+}
+
+/** Map one non-ranked hardware's coverage and read to its not-ranked reason. */
+function notRankedReasonFor(status: OverviewHardwareStatus): OverviewNotRankedReason {
+  switch (status.coverage.kind) {
+    case 'standard_only': {
+      return 'standard_decode_only';
+    }
+    case 'alternate_precision_only': {
+      return 'other_precision_only';
+    }
+    case 'unsupported_precision_only': {
+      return 'int4_bf16_only';
+    }
+    case 'other_engine_group_only': {
+      return 'different_serving_cohort';
+    }
+    case 'no_workload_data': {
+      return 'no_8k1k_data';
+    }
+    case 'comparable_spec': {
+      // Comparable-spec hardware reaches here only without an exact @50 read.
+      // Classify by INTERACTIVITY direction, not the boundary name: 'unreachable'
+      // = tier 50 above the frontier's max, i.e. even the fastest measured point
+      // is < 50 tok/s/user (the config is too slow, ever) → cannot_reach_at50.
+      // 'clamped_low' = tier below the frontier's min, i.e. every measured point
+      // is > 50 tok/s/user — an under-swept sweep gap more concurrency would
+      // cross, NOT incapability → no_exact_at50.
+      return status.primary.boundary === 'unreachable' ? 'cannot_reach_at50' : 'no_exact_at50';
+    }
+  }
+}
+
+/**
+ * One not-ranked entry per page hardware with no visible value — no exact @50
+ * read in the primary block and no ranked read in a `ranked` secondary. Reads
+ * the same coverage/status machinery the blocks use, so it is a derived view
+ * rather than a parallel source of truth.
+ */
+function buildNotRanked(
+  primaryStatuses: readonly OverviewHardwareStatus[],
+  secondary: OverviewSecondaryPrecision | null,
+): OverviewNotRankedEntry[] {
+  const secondaryRanked =
+    secondary?.state === 'ranked'
+      ? new Set(
+          secondary.hardwareStatuses
+            .filter((status) => isExactTierRead(status.primary))
+            .map((status) => status.hardware),
+        )
+      : new Set<string>();
+
+  const entries: OverviewNotRankedEntry[] = [];
+  for (const status of primaryStatuses) {
+    if (isExactTierRead(status.primary) || secondaryRanked.has(status.hardware)) continue;
+    entries.push({
+      hardware: status.hardware,
+      hardwareLabel: status.hardwareLabel,
+      reason: notRankedReasonFor(status),
+      precisions: status.coverage.availablePrecisions.filter((p) =>
+        OVERVIEW_PRECISIONS.includes(p),
+      ),
+    });
+  }
+  return entries;
+}
+
+function emptyModelSummary(
+  model: Model,
+  emptyReason: NonNullable<OverviewModelSummary['emptyReason']>,
+  order: readonly OverviewHardwareOrderEntry[],
+  workloadRows: readonly BenchmarkRow[],
+): OverviewModelSummary {
+  const hardwareStatuses = buildHardwareStatuses(
+    buildCoverage(order, workloadRows, null, new Set()),
+    [],
+    false,
+  );
+  return {
+    model,
+    modelLabel: getModelLabel(model),
+    selectedPrecision: null,
+    hardwareStatuses,
+    comparisonGroups: [],
+    secondary: null,
+    notRanked: buildNotRanked(hardwareStatuses, null),
+    latestWorkloadDate: latestWorkloadDateOf(workloadRows),
+    emptyReason,
+  };
+}
+
+export function buildOverviewModelSummary(
+  model: Model,
+  rows: BenchmarkRow[],
+  hardwareOrder: readonly OverviewHardwareOrderEntry[] = buildOverviewHardwareOrder(rows),
+): OverviewModelSummary {
+  const workloadRows = overviewWorkloadRows(rows);
+  if (workloadRows.length === 0) {
+    return emptyModelSummary(model, 'no_8k1k_single_turn_data', hardwareOrder, workloadRows);
+  }
+
+  if (!workloadRows.some((row) => OVERVIEW_PRECISIONS.includes(row.precision))) {
+    return emptyModelSummary(model, 'no_fp4_fp8_data', hardwareOrder, workloadRows);
+  }
+
+  // FP4 and FP8 speculative configs, built once and reused for selection, the
+  // primary block and the secondary block — the two precisions never blend.
+  const configsByPrecision = new Map(
+    OVERVIEW_PRECISIONS.map(
+      (precision) => [precision, buildPrecisionConfigs(model, workloadRows, precision)] as const,
+    ),
+  );
+  const selectedPrecision = selectPrimaryFrom(configsByPrecision);
+  const secondaryPrecision = OVERVIEW_PRECISIONS.find((p) => p !== selectedPrecision)!;
+  const primaryConfigs = configsByPrecision.get(selectedPrecision)!;
+
+  const primary = buildPrecisionBlock(
+    model,
+    selectedPrecision,
+    primaryConfigs,
+    hardwareOrder,
+    workloadRows,
+  );
+  const secondary = buildSecondaryPrecision(
+    model,
+    secondaryPrecision,
+    configsByPrecision.get(secondaryPrecision)!,
+    exactPrimaryHardware(primaryConfigs),
+    hardwareOrder,
+    workloadRows,
+  );
+
+  return {
+    model,
+    modelLabel: getModelLabel(model),
+    selectedPrecision,
+    hardwareStatuses: primary.hardwareStatuses,
+    comparisonGroups: primary.comparisonGroups,
+    secondary,
+    notRanked: buildNotRanked(primary.hardwareStatuses, secondary),
+    latestWorkloadDate: latestWorkloadDateOf(workloadRows),
+    emptyReason: null,
+  };
+}
+
+/**
+ * Assemble the whole page from rows already grouped by DISPLAY model. Iterating
+ * DEFAULT_MODELS fixes the output order and renders every active model — a model
+ * with no rows falls through to its coverage-only states, exactly as the live
+ * path does. The page-wide hardware order and dataset date come from the union
+ * of every model's rows. The live server path and the e2e fixture path feed this
+ * same function; only the row source differs.
+ */
+export function assembleOverviewPageData(
+  rowsByModel: Record<string, BenchmarkRow[]>,
+): OverviewPageData {
+  const perModel = [...DEFAULT_MODELS].map((model) => ({ model, rows: rowsByModel[model] ?? [] }));
+  const allRows = perModel.flatMap(({ rows }) => rows);
+  const hardwareOrder = buildOverviewHardwareOrder(allRows);
+  return {
+    models: perModel.map(({ model, rows }) =>
+      buildOverviewModelSummary(model, rows, hardwareOrder),
+    ),
+    hardwareOrder,
+    datasetThroughDate: overviewDatasetThroughDate(allRows),
+  };
+}
diff --git a/packages/app/src/lib/overview-links.test.ts b/packages/app/src/lib/overview-links.test.ts
new file mode 100644
index 00000000..0f6c4f35
--- /dev/null
+++ b/packages/app/src/lib/overview-links.test.ts
@@ -0,0 +1,100 @@
+import { describe, expect, it } from 'vitest';
+
+import { Model, Precision } from './data-mappings';
+import type { OverviewConfigResult, OverviewModelSummary } from './overview-data';
+import { buildOverviewDashboardHref } from './overview-links';
+
+const RUN_URL = 'https://github.com/SemiAnalysisAI/InferenceX/actions/runs/26714221123';
+
+/** Query the default fixture produces: one source run, so the run is pinned. */
+const PINNED_QUERY =
+  'g_model=Qwen-3.5-397B-A17B&g_rundate=2026-07-18&g_runid=26714221123&i_seq=8k%2F1k' +
+  '&i_prec=fp4&i_gpus=b200_sglang_mtp&i_spec=mtp&i_disagg=agg&i_optimal=1&i_advlabel=1';
+
+function config(overrides: Partial<OverviewConfigResult> = {}): OverviewConfigResult {
+  return {
+    key: 'qwen3.5|b200|sglang|mtp|agg|fp4',
+    dbModel: 'qwen3.5',
+    hardware: 'b200',
+    hardwareLabel: 'B200',
+    hwKey: 'b200_sglang_mtp',
+    framework: 'sglang',
+    frameworkLabel: 'SGLang',
+    specMethod: 'mtp',
+    specLabel: 'MTP',
+    disagg: false,
+    precision: Precision.FP4,
+    offloadMode: 'off',
+    isMultinode: false,
+    numPrefillGpu: 8,
+    numDecodeGpu: 8,
+    totalGpu: 8,
+    parallelism: 'TP8',
+    image: 'inferencex/sglang:v0.5.0',
+    sourceRunUrls: [RUN_URL],
+    tierValues: [{ tier: 50, value: 1000, boundary: 'interpolated', evidenceDate: null }],
+    latestDate: '2026-07-18',
+    oldestFrontierDate: '2026-07-11',
+    ...overrides,
+  };
+}
+
+function summary(overrides: Partial<OverviewModelSummary> = {}): OverviewModelSummary {
+  return {
+    model: Model.Qwen3_5,
+    modelLabel: 'Qwen 3.5',
+    selectedPrecision: Precision.FP4,
+    hardwareStatuses: [],
+    comparisonGroups: [],
+    secondary: null,
+    notRanked: [],
+    latestWorkloadDate: null,
+    emptyReason: null,
+    ...overrides,
+  };
+}
+
+describe('buildOverviewDashboardHref', () => {
+  it('pins model, run, workload and exact configuration on the English route', () => {
+    expect(buildOverviewDashboardHref('en', summary(), config())).toBe(
+      `/inference?${PINNED_QUERY}`,
+    );
+  });
+
+  it('selects the disaggregated deployment mode for a disaggregated configuration', () => {
+    const href = buildOverviewDashboardHref(
+      'en',
+      summary(),
+      config({ disagg: true, hwKey: 'gb200_dynamo-trt-disagg_mtp' }),
+    );
+
+    expect(href).toBe(
+      '/inference?g_model=Qwen-3.5-397B-A17B&g_rundate=2026-07-18&g_runid=26714221123' +
+        '&i_seq=8k%2F1k&i_prec=fp4&i_gpus=gb200_dynamo-trt-disagg_mtp&i_spec=mtp' +
+        '&i_disagg=disagg&i_optimal=1&i_advlabel=1',
+    );
+  });
+
+  it('writes g_model even when it equals the dashboard default model', () => {
+    const href = buildOverviewDashboardHref(
+      'en',
+      summary({ model: Model.DeepSeek_V4_Pro }),
+      config({ precision: Precision.FP8 }),
+    );
+
+    expect(href).toContain('g_model=DeepSeek-V4-Pro');
+    expect(href).toContain('i_prec=fp8');
+  });
+
+  it('maps specMethod to the dashboard mtp/stp filter bucket, not the raw DB value', () => {
+    expect(buildOverviewDashboardHref('en', summary(), config({ specMethod: 'eagle' }))).toContain(
+      'i_spec=mtp',
+    );
+    expect(buildOverviewDashboardHref('en', summary(), config({ specMethod: 'none' }))).toContain(
+      'i_spec=stp',
+    );
+    expect(buildOverviewDashboardHref('en', summary(), config({ specMethod: 'mtp' }))).toContain(
+      'i_spec=mtp',
+    );
+  });
+});
diff --git a/packages/app/src/lib/overview-links.ts b/packages/app/src/lib/overview-links.ts
new file mode 100644
index 00000000..8e9deb5e
--- /dev/null
+++ b/packages/app/src/lib/overview-links.ts
@@ -0,0 +1,89 @@
+/**
+ * @file overview-links.ts
+ * @description Evidence links for the overview scorecard: a pre-filtered
+ * /inference view and the CI run a configuration came from. Pure and
+ * React-free so the server-rendered overview page can build them.
+ */
+
+import { runIdFromRunUrl } from './known-issues';
+import type { OverviewConfigResult, OverviewModelSummary } from './overview-data';
+import type { UrlStateParams } from './url-state';
+
+/** The single-turn 8K-in/1K-out workload every overview link filters to. */
+const OVERVIEW_WORKLOAD_SEQ = '8k/1k';
+
+/** The `/inference` route base for a locale — shared by every overview link. */
+function inferenceRoute(locale: 'en' | 'zh'): string {
+  return locale === 'zh' ? '/zh/inference' : '/inference';
+}
+
+/**
+ * Maps a raw DB `spec_method` to the dashboard's `SpecMode` filter bucket
+ * (mirrors `pointSpecMode` in quickFilters.ts, minus its hwKey suffix check —
+ * overview `specMethod` comes straight from `spec_method`).
+ */
+function dashboardSpecMode(specMethod: string): 'mtp' | 'stp' {
+  return specMethod === 'none' || specMethod === '' ? 'stp' : 'mtp';
+}
+
+/**
+ * The one run backing a configuration, or null when it has none, has several,
+ * or its single source URL names no run (a run list rather than a run). Both
+ * helpers below read this one predicate, so the `g_runid` pin and the source-run
+ * link can never disagree about whether a single run backs the configuration.
+ */
+function soleSourceRun(config: OverviewConfigResult): { url: string; id: string } | null {
+  if (config.sourceRunUrls.length !== 1) return null;
+  const url = config.sourceRunUrls[0];
+  const id = runIdFromRunUrl(url);
+  return id === null ? null : { url, id };
+}
+
+/**
+ * Inference-dashboard link narrowed to the configuration the overview ranked:
+ * its model, run date, workload, precision, hardware/framework/spec key and
+ * deployment mode. The run is pinned only when a single run produced the
+ * configuration — pinning one of several would hide the rest of its frontier.
+ *
+ * This is a filtered view, not a proof of topology: `i_gpus` selects a
+ * hardware/framework/spec key, which can still hold more than one parallelism
+ * or GPU-count topology.
+ */
+export function buildOverviewDashboardHref(
+  locale: 'en' | 'zh',
+  model: OverviewModelSummary,
+  config: OverviewConfigResult,
+): string {
+  const params: UrlStateParams = {
+    g_model: model.model,
+    g_rundate: config.latestDate,
+    g_runid: soleSourceRun(config)?.id,
+    i_seq: OVERVIEW_WORKLOAD_SEQ,
+    i_prec: config.precision,
+    i_gpus: config.hwKey,
+    i_spec: dashboardSpecMode(config.specMethod),
+    i_disagg: config.disagg ? 'disagg' : 'agg',
+    i_optimal: '1',
+    i_advlabel: '1',
+  };
+
+  const query = new URLSearchParams();
+  for (const [key, value] of Object.entries(params)) {
+    if (value !== undefined) query.set(key, value);
+  }
+  return `${inferenceRoute(locale)}?${query}`;
+}
+
+/**
+ * Model-level dashboard view: the workload and precision this page ranked, with
+ * no configuration pinned. Cohort-level evidence links narrow it further.
+ */
+export function detailHref(locale: 'en' | 'zh', model: OverviewModelSummary): string {
+  const query = new URLSearchParams({
+    g_model: model.model,
+    i_seq: OVERVIEW_WORKLOAD_SEQ,
+    ...(model.selectedPrecision ? { i_prec: model.selectedPrecision } : {}),
+    i_optimal: '1',
+  });
+  return `${inferenceRoute(locale)}?${query}`;
+}
diff --git a/packages/app/src/lib/tco-feed.ts b/packages/app/src/lib/tco-feed.ts
index 12630d4e..9041b130 100644
--- a/packages/app/src/lib/tco-feed.ts
+++ b/packages/app/src/lib/tco-feed.ts
@@ -61,6 +61,14 @@ export interface TcoFeedRow {
   latest_date: string;
   /** Oldest benchmark date among the frontier knots (staleness flag). */
   oldest_frontier_date: string;
+  /**
+   * Run dates of the frontier knot(s) actually backing THIS tier read: the
+   * bracketing pair when interpolated, the single min-interactivity knot when
+   * clamped_low, null when unreachable. Unlike latest/oldest_frontier_date
+   * (whole-frontier freshness), it never blends knots outside the tier, so a
+   * downstream result can carry its own evidence date. Ordered ascending.
+   */
+  evidence_date: { from: string; to: string } | null;
 }
 
 export const DEFAULT_TIERS: readonly number[] = [30, 50, 75, 100];
@@ -197,6 +205,27 @@ interface FrontierPoint {
 
 const round3 = (v: number): number => Math.round(v * 1000) / 1000;
 
+/** Two knots' dates as an ascending {from,to} evidence range. */
+function evidenceRange(a: FrontierPoint, b: FrontierPoint): { from: string; to: string } {
+  return a.date <= b.date ? { from: a.date, to: b.date } : { from: b.date, to: a.date };
+}
+
+/**
+ * The frontier knots backing an interpolated read at `tier`: the bracketing
+ * pair xs[lo] ≤ tier ≤ xs[lo+1], collapsing to a single knot when `tier` lands
+ * on an endpoint. `frontier` is sorted ascending by interactivity.
+ */
+function bracketKnots(frontier: FrontierPoint[], tier: number): [FrontierPoint, FrontierPoint] {
+  const last = frontier.length - 1;
+  if (tier <= frontier[0].interactivity) return [frontier[0], frontier[0]];
+  if (tier >= frontier[last].interactivity) return [frontier[last], frontier[last]];
+  let lo = 0;
+  for (let i = 0; i < last; i++) {
+    if (frontier[i].interactivity <= tier) lo = i;
+  }
+  return [frontier[lo], frontier[lo + 1]];
+}
+
 /**
  * Compute the feed: for each (workload, hardware), build the
  * (interactivity, output tok/s/GPU) Pareto frontier across every config —
@@ -268,16 +297,22 @@ export function computeTcoFeed(
       for (const tier of tiers) {
         let value: number;
         let boundary: TcoTierBoundary;
+        let evidenceDate: { from: string; to: string } | null;
         if (tier > maxIv) {
           value = 0;
           boundary = 'unreachable';
+          evidenceDate = null;
         } else if (tier < minIv) {
           value = ys[0];
           boundary = 'clamped_low';
+          // The clamped read is the min-interactivity knot's throughput.
+          evidenceDate = evidenceRange(frontier[0], frontier[0]);
         } else {
           const raw = hermiteInterpolate(xs, ys, slopes, tier);
           value = Math.max(yLo, Math.min(yHi, raw));
           boundary = 'interpolated';
+          const [lo, hi] = bracketKnots(frontier, tier);
+          evidenceDate = evidenceRange(lo, hi);
         }
         out.push({
           hardware,
@@ -290,6 +325,7 @@ export function computeTcoFeed(
           frontier_max_interactivity: round3(maxIv),
           latest_date: latest,
           oldest_frontier_date: oldest,
+          evidence_date: evidenceDate,
         });
       }
     }
diff --git a/packages/app/src/lib/toggle-set.ts b/packages/app/src/lib/toggle-set.ts
new file mode 100644
index 00000000..963d3ea4
--- /dev/null
+++ b/packages/app/src/lib/toggle-set.ts
@@ -0,0 +1,32 @@
+/**
+ * Pure toggle state transition.
+ *
+ * - If all items are active and one is toggled, only that item remains active (solo).
+ * - If only one item is active and it's toggled, all items become active (restore all).
+ * - Otherwise, the item is simply added or removed from the active set.
+ *
+ * Kept free of React so server-only modules (`exclusion.ts`, and through it the
+ * overview data layer) can reuse it without pulling client hooks into a Server
+ * Component graph.
+ */
+export function computeToggle(prev: Set<string>, item: string, allItems: Set<string>): Set<string> {
+  const allAreActive = prev.size === allItems.size;
+  const isActive = prev.has(item);
+
+  if (isActive) {
+    if (allAreActive) {
+      // Solo: only keep the clicked item
+      return new Set([item]);
+    } else if (prev.size === 1) {
+      // Restore all: re-enable everything
+      return allItems;
+    }
+    // Remove the clicked item
+    const next = new Set(prev);
+    next.delete(item);
+    return next;
+  }
+  // Add the clicked item
+  const next = new Set([...prev, item]);
+  return next;
+}

From 1febe16bdc04899c6b3f7206f92130ee401be716 Mon Sep 17 00:00:00 2001
From: Wenyao Gao <wgao11@u.rochester.edu>
Date: Wed, 22 Jul 2026 14:23:05 -0700
Subject: [PATCH 02/13] feat(overview): infinity placeholder chips for missing
 results

---
 packages/app/cypress/e2e/overview.cy.ts       | 16 +++++++---
 .../overview/overview-scorecard.tsx           | 32 ++++++++++++++-----
 2 files changed, 35 insertions(+), 13 deletions(-)

diff --git a/packages/app/cypress/e2e/overview.cy.ts b/packages/app/cypress/e2e/overview.cy.ts
index bdd92f69..1b258f26 100644
--- a/packages/app/cypress/e2e/overview.cy.ts
+++ b/packages/app/cypress/e2e/overview.cy.ts
@@ -89,11 +89,12 @@ describe('Overview page', () => {
       cy.contains('Jul 18').should('exist');
       cy.contains('381').should('exist');
       cy.contains('Only exact result').should('exist');
-      // Every remaining hardware carries one reason, both clamp directions included.
-      cy.contains('cannot reach @50').should('exist');
-      cy.contains('no exact @50 result').should('exist');
-      cy.contains('no 8K/1K data').should('exist');
-      cy.contains('standard decode only').should('exist');
+      // Every remaining hardware carries an ∞ chip whose title/aria holds the
+      // reason, both clamp directions included; ∞ never renders with a percent.
+      cy.get('[title="cannot reach @50"]').should('exist').and('contain', '∞');
+      cy.get('[title="no exact @50 result"]').should('exist').and('contain', '∞');
+      cy.get('[title="no 8K/1K data"]').should('exist').and('contain', '∞');
+      cy.get('[title="standard decode only"]').should('exist').and('contain', '∞');
       // Each ranked value links into its own pre-filtered dashboard view.
       cy.get('a[href*="i_gpus="]')
         .first()
@@ -102,6 +103,11 @@ describe('Overview page', () => {
         .and('include', 'i_gpus=b300_sglang_mtp')
         .and('include', 'i_spec=mtp');
     });
+    // The legend lives once in the page footnotes; ∞ never renders as a percent.
+    cy.contains('∞ = no comparable result').should('exist');
+    cy.get('body')
+      .invoke('text')
+      .should('not.match', /∞\s*%/);
   });
 
   it('opens a ranked secondary only when the other precision adds comparable hardware', () => {
diff --git a/packages/app/src/components/overview/overview-scorecard.tsx b/packages/app/src/components/overview/overview-scorecard.tsx
index a1c02907..f80c1d07 100644
--- a/packages/app/src/components/overview/overview-scorecard.tsx
+++ b/packages/app/src/components/overview/overview-scorecard.tsx
@@ -65,7 +65,7 @@ export const OVERVIEW_STRINGS = {
       `Open filtered dashboard: ${model} · ${config}`,
     evidenceHighAria: (model: string, config: string) =>
       `Open filtered dashboard @${OVERVIEW_HIGH_TIER} tok/s/user leader: ${model} · ${config}`,
-    notRankedLine: (entries: string) => `Not ranked: ${entries}`,
+    infinityLegend: '∞ = no comparable result',
     notRankedReasons: {
       standard_decode_only: 'standard decode only',
       int4_bf16_only: 'INT4/BF16 only',
@@ -125,7 +125,7 @@ export const OVERVIEW_STRINGS = {
       `打开筛选后的仪表板：${model} · ${config}`,
     evidenceHighAria: (model: string, config: string) =>
       `打开筛选后的仪表板 @${OVERVIEW_HIGH_TIER} tok/s/user：${model} · ${config}`,
-    notRankedLine: (entries: string) => `未参与排名：${entries}`,
+    infinityLegend: '∞ = 无可比结果',
     notRankedReasons: {
       standard_decode_only: '仅标准解码',
       int4_bf16_only: '仅 INT4/BF16',
@@ -551,12 +551,27 @@ function NotRankedLine({
   // coverage 'measured' list — is not repeated here; dedup against the secondary's
   // measured labels so each hardware surfaces in exactly one place.
   const secondaryMeasured = new Set(model.secondary?.measuredHardware);
-  const entries = model.notRanked
-    .filter((entry) => !secondaryMeasured.has(entry.hardwareLabel))
-    .map((entry) => `${entry.hardwareLabel} — ${notRankedReason(entry, strings)}`)
-    .join(' · ');
-  if (entries === '') return null;
-  return <p className="text-xs text-muted-foreground">{strings.notRankedLine(entries)}</p>;
+  const entries = model.notRanked.filter((entry) => !secondaryMeasured.has(entry.hardwareLabel));
+  if (entries.length === 0) return null;
+  // ∞ marks missing/unavailable only — it never enters ranking or gap math, so it
+  // must never render with a percent. The reason rides in title/aria; the page
+  // footer carries the legend.
+  return (
+    <p className="flex flex-wrap gap-x-3 gap-y-0.5 text-xs text-muted-foreground">
+      {entries.map((entry) => {
+        const reason = notRankedReason(entry, strings);
+        return (
+          <span
+            key={entry.hardware}
+            title={reason}
+            aria-label={`${entry.hardwareLabel}: ${reason}`}
+          >
+            {entry.hardwareLabel} <span aria-hidden="true">∞</span>
+          </span>
+        );
+      })}
+    </p>
+  );
 }
 
 /** The full per-model hierarchy, identical on both surfaces. */
@@ -729,6 +744,7 @@ export function OverviewMethodology({ strings }: { strings: OverviewStrings }) {
   return (
     <div className="space-y-1 border-t border-border/50 px-4 py-3 text-xs leading-snug text-muted-foreground lg:px-6">
       <p>{strings.methodologyNote}</p>
+      <p>{strings.infinityLegend}</p>
       <p>{strings.cohortNote}</p>
       <p>{strings.interpolationNote}</p>
     </div>

From 76af0922a528684161ab9d0e009ddfe976deeff9 Mon Sep 17 00:00:00 2001
From: Wenyao Gao <wgao11@u.rochester.edu>
Date: Wed, 22 Jul 2026 18:33:18 -0700
Subject: [PATCH 03/13] feat(overview): per-platform matrix with tier switcher
 and stack labels

---
 packages/app/cypress/e2e/overview.cy.ts       | 335 +++++--
 .../cypress/fixtures/api/overview-rows.json   |  39 +-
 packages/app/src/app/overview/page.tsx        |  12 +-
 packages/app/src/app/zh/overview/page.tsx     |  12 +-
 .../src/components/overview/overview-page.tsx |   4 +-
 .../overview/overview-scorecard.tsx           | 842 ++++++----------
 packages/app/src/lib/overview-data.server.ts  |  18 +-
 packages/app/src/lib/overview-data.test.ts    | 871 +++++++----------
 packages/app/src/lib/overview-data.ts         | 905 +++++-------------
 packages/app/src/lib/overview-links.test.ts   |  18 +-
 packages/app/src/lib/overview-links.ts        |  22 +-
 packages/app/src/lib/tco-feed.test.ts         |  28 +
 packages/app/src/lib/tco-feed.ts              |   1 +
 13 files changed, 1239 insertions(+), 1868 deletions(-)

diff --git a/packages/app/cypress/e2e/overview.cy.ts b/packages/app/cypress/e2e/overview.cy.ts
index 1b258f26..89fb5039 100644
--- a/packages/app/cypress/e2e/overview.cy.ts
+++ b/packages/app/cypress/e2e/overview.cy.ts
@@ -1,11 +1,3 @@
-// Focused smoke coverage for /overview. In fixtures mode the page is served by
-// the synthetic cypress/fixtures/api/overview-rows.json through the real data
-// builder, so every expected value below was derived by running the assembler
-// over that fixture (the drift guard in overview-data.test.ts locks the same
-// values) — never eyeballed. The fixture exercises every 8.7 state: coverage-driven
-// primary precision, ranked vs coverage secondary, all not-ranked reasons, and —
-// uniquely — a cross-day evidence range the live dataset does not produce.
-
 const MODEL_LABELS = [
   'DeepSeek V4 Pro 1.6T',
   'Kimi K2.5/2.6/2.7-Code 1T',
@@ -14,20 +6,22 @@ const MODEL_LABELS = [
   'Qwen3.5 397B',
 ];
 
-/** The page must never scroll sideways: the whole comparison has to fit. */
+const PLATFORM_HEADERS = [
+  'Model',
+  'B200',
+  'MI355X',
+  'B300',
+  'GB200 NVL72',
+  'GB300 NVL72',
+  'Details',
+];
+
 function expectNoHorizontalOverflow() {
   cy.document().then((doc) => {
     expect(doc.documentElement.scrollWidth).to.be.lte(doc.documentElement.clientWidth);
   });
 }
 
-/**
- * A narrow document is not enough: a table parked in an `overflow-x` region
- * keeps the document narrow while still scrolling sideways, so nothing inside
- * the surface may be wider than its own box either. `sr-only` clips (1px) and
- * non-replaced inline boxes are exempt — CSSOM reports a spurious `scrollWidth`
- * for inline boxes in Firefox, and `overflow` cannot apply to them anyway.
- */
 function expectNoHorizontalScroller(testId: string) {
   cy.get(`[data-testid="${testId}"]`).then(([surface]) => {
     const scrollers = [surface, ...surface.querySelectorAll('*')]
@@ -46,112 +40,275 @@ function desktopModel(model: string) {
   return cy.get(`[data-testid="overview-desktop-model"][data-model="${model}"]`);
 }
 
+function mobileModel(model: string) {
+  return cy.get(`[data-testid="overview-mobile-model"][data-model="${model}"]`);
+}
+
+function pair(pairId: string) {
+  return cy.get(`[data-testid="overview-pair"][data-pair="${pairId}"]`);
+}
+
 describe('Overview page', () => {
-  it('summarizes every active model without interactive widgets', () => {
-    cy.viewport(1280, 800);
+  it('renders the full platform matrix for every active model', () => {
+    cy.viewport(1280, 900);
     cy.visit('/overview');
 
     cy.contains('h1', 'AI Inference Overview').should('exist');
-    cy.get('[data-testid="overview-desktop-matrix"]').should('be.visible');
-    cy.get('[data-testid="overview-desktop-matrix"] h2').should('have.length', MODEL_LABELS.length);
+    cy.contains(
+      'Every active model across MI355X, B200, B300, GB200 and GB300 at a glance.',
+    ).should('exist');
+    cy.contains('Best validated stack per platform').should('exist');
+    cy.contains('Database snapshot through Jul 18').should('exist');
+    cy.get('[data-testid="overview-desktop-matrix"]')
+      .should('be.visible')
+      .within(() => {
+        cy.get('thead th').then(($headers) => {
+          expect([...$headers].map((header) => header.textContent?.trim())).to.deep.equal(
+            PLATFORM_HEADERS,
+          );
+        });
+        cy.get('[data-testid="overview-desktop-model"]').should('have.length', MODEL_LABELS.length);
+        cy.get('[data-testid="overview-baseline"]').should('have.length', MODEL_LABELS.length);
+        cy.get('[data-testid="overview-pair"]').should('have.length', MODEL_LABELS.length * 4);
+        cy.get('details, summary, button').should('not.exist');
+        cy.contains(/PRIMARY|Ranked results/).should('not.exist');
+      });
     for (const label of MODEL_LABELS) {
       cy.get('[data-testid="overview-desktop-matrix"]').should('contain.text', label);
     }
-    // Database-wide freshness line + the methodology footnote's once-per-page note.
-    cy.contains('Database snapshot through Jul 18').should('exist');
-    cy.contains(
-      'The default precision maximizes comparable hardware coverage. Not ranked does not mean slower.',
-    ).should('exist');
-    // A one-glance summary, not an interactive widget.
-    cy.get('[data-testid="overview-desktop-matrix"]').within(() => {
-      cy.get('details, summary, button').should('not.exist');
+  });
+
+  it('shows per-cell best reads with precision badges, dates, links, and matched deltas', () => {
+    cy.viewport(1280, 900);
+    cy.visit('/overview');
+
+    desktopModel('Qwen-3.5-397B-A17B').within(() => {
+      cy.get('[data-testid="overview-baseline"][data-hardware="b200"]')
+        .should('contain.text', '900')
+        .and('contain.text', 'FP8');
+      pair('mi355x-vs-b200').within(() => {
+        cy.contains('SGLang · FP8').should('exist');
+        cy.get('[data-testid="overview-pair-value"][data-hardware="mi355x"]')
+          .should('contain.text', '760')
+          .find('a')
+          .should('have.attr', 'title', 'MI355X · SGLang · FP8 · MTP');
+        cy.get('[data-testid="overview-pair-value"][data-hardware="mi355x"] a')
+          .should('have.attr', 'href')
+          .and('include', 'g_model=Qwen-3.5-397B-A17B')
+          .and('include', 'i_prec=fp8')
+          .and('include', 'i_gpus=mi355x_sglang_mtp');
+        cy.get('[data-testid="overview-pair-evidence-date"][data-hardware="mi355x"]').should(
+          'have.text',
+          'Jul 18',
+        );
+        cy.get('[data-testid="overview-pair-delta"]').should('have.text', '−16% vs B200');
+        cy.contains('At 100, MI355X leads').should('exist');
+      });
+    });
+
+    desktopModel('DeepSeek-V4-Pro').within(() => {
+      cy.get('[data-testid="overview-baseline"][data-hardware="b200"]')
+        .should('contain.text', '900')
+        .and('contain.text', 'FP4');
+      pair('b300-vs-b200').within(() => {
+        cy.get('[data-testid="overview-pair-value"][data-hardware="b300"]').should(
+          'contain.text',
+          '1,122',
+        );
+        cy.get('[data-testid="overview-pair-evidence-date"][data-hardware="b300"]').should(
+          'have.text',
+          'Jun 24–Jul 4',
+        );
+        cy.get('[data-testid="overview-pair-delta"]').should('have.text', '+25% vs B200');
+      });
+    });
+  });
+
+  it('never mixes precisions or releases in a delta and says so instead', () => {
+    cy.viewport(1280, 900);
+    cy.visit('/overview');
+
+    desktopModel('DeepSeek-V4-Pro').within(() => {
+      pair('gb200-vs-b200').within(() => {
+        cy.get('[data-testid="overview-pair-value"][data-hardware="gb200"]')
+          .should('contain.text', '600')
+          .parents('[data-testid="overview-pair"]')
+          .should('contain.text', 'FP8');
+        cy.get('[data-testid="overview-pair-mismatch"]').should(
+          'have.text',
+          'FP8 vs FP4 · no comparable delta',
+        );
+        cy.get('[data-testid="overview-pair-delta"]').should('not.exist');
+      });
+    });
+
+    desktopModel('Qwen-3.5-397B-A17B').within(() => {
+      pair('b300-vs-b200').within(() => {
+        cy.get('[data-testid="overview-pair-value"][data-hardware="b300"]').should(
+          'contain.text',
+          '1,151',
+        );
+        cy.get('[data-testid="overview-pair-mismatch"]').should(
+          'have.text',
+          'FP4 vs FP8 · no comparable delta',
+        );
+        cy.get('[data-testid="overview-pair-delta"]').should('not.exist');
+      });
     });
   });
 
-  it('ranks the primary precision, dates each read, and accounts for every hardware', () => {
-    cy.viewport(1280, 800);
+  it('distinguishes candidate, baseline, and whole-row missing results without a percent', () => {
+    cy.viewport(1280, 900);
     cy.visit('/overview');
 
     desktopModel('DeepSeek-V4-Pro').within(() => {
-      // Primary precision, its leader, and the runner-up's own signed delta (ASCII minus).
-      cy.contains('FP4 · PRIMARY @50').should('exist');
-      cy.contains('B300').should('exist');
-      cy.contains('1,122').should('exist');
-      cy.contains('Leader').should('exist');
-      cy.contains('-20%').should('exist');
-      // FP8 measured one platform → one compact coverage line, not a second table.
-      cy.contains('FP8 coverage: GB200 NVL72 measured; insufficient comparable results.').should(
-        'exist',
-      );
-      // Leader @50 is bracketed by two run days (en-dash range); same-day reads and
-      // the @100 read (its own leader) collapse to a single date.
-      cy.contains('Jun 24–Jul 4').should('exist');
-      cy.contains('Jul 18').should('exist');
-      cy.contains('381').should('exist');
-      cy.contains('Only exact result').should('exist');
-      // Every remaining hardware carries an ∞ chip whose title/aria holds the
-      // reason, both clamp directions included; ∞ never renders with a percent.
-      cy.get('[title="cannot reach @50"]').should('exist').and('contain', '∞');
-      cy.get('[title="no exact @50 result"]').should('exist').and('contain', '∞');
-      cy.get('[title="no 8K/1K data"]').should('exist').and('contain', '∞');
-      cy.get('[title="standard decode only"]').should('exist').and('contain', '∞');
-      // Each ranked value links into its own pre-filtered dashboard view.
-      cy.get('a[href*="i_gpus="]')
-        .first()
-        .should('have.attr', 'href')
-        .and('include', 'g_model=DeepSeek-V4-Pro')
-        .and('include', 'i_gpus=b300_sglang_mtp')
-        .and('include', 'i_spec=mtp');
+      pair('mi355x-vs-b200').within(() => {
+        cy.get('[data-testid="overview-pair-missing"][data-hardware="mi355x"]')
+          .should('contain.text', '∞')
+          .and('have.attr', 'title', 'no exact @50 result');
+      });
+      pair('gb300-vs-b200').within(() => {
+        cy.get('[data-testid="overview-pair-missing"][data-hardware="gb300"]')
+          .should('contain.text', '∞')
+          .and('have.attr', 'title', 'cannot reach @50');
+      });
+    });
+
+    desktopModel('MiniMax-M3').within(() => {
+      cy.get('[data-testid="overview-baseline"] [data-testid="overview-pair-missing"]')
+        .should('contain.text', '∞')
+        .and('have.attr', 'title', 'no 8K/1K data');
+      pair('gb300-vs-b200').within(() => {
+        cy.get('[data-testid="overview-pair-value"][data-hardware="gb300"]').should(
+          'contain.text',
+          '700',
+        );
+        cy.get('[data-testid="overview-pair-delta"]').should('not.exist');
+        cy.get('[data-testid="overview-pair-mismatch"]').should('not.exist');
+      });
+    });
+
+    desktopModel('Kimi-K2.5').within(() => {
+      cy.get('[data-testid="overview-pair-missing"]').should('have.length', 5);
+      cy.get('[data-testid="overview-pair-value"]').should('not.exist');
     });
-    // The legend lives once in the page footnotes; ∞ never renders as a percent.
     cy.contains('∞ = no comparable result').should('exist');
     cy.get('body')
       .invoke('text')
       .should('not.match', /∞\s*%/);
   });
 
-  it('opens a ranked secondary only when the other precision adds comparable hardware', () => {
-    cy.viewport(1280, 800);
+  it('re-renders the whole matrix at the service level the URL names, via plain links', () => {
+    cy.viewport(1280, 900);
     cy.visit('/overview');
 
-    // Qwen: FP8 ranks a comparable pair AND adds MI355X, so it renders subordinate rows.
+    cy.get('[data-testid="overview-tier-switcher"]').within(() => {
+      cy.get('[aria-current="page"]').should('have.text', '50');
+      cy.get('a').should('have.length', 3);
+      cy.contains('a', '30').should('have.attr', 'href', '/overview?tier=30');
+      cy.contains('a', '100').should('have.attr', 'href', '/overview?tier=100').click();
+    });
+
+    cy.location('search').should('eq', '?tier=100');
+    cy.contains('Output tok/s/GPU @100 tok/s/user').should('exist');
+    cy.get('[data-testid="overview-tier-switcher"]').within(() => {
+      cy.get('[aria-current="page"]').should('have.text', '100');
+      cy.contains('a', '50').should('have.attr', 'href', '/overview');
+    });
+
     desktopModel('Qwen-3.5-397B-A17B').within(() => {
-      cy.contains('FP4 · PRIMARY @50').should('exist');
-      cy.contains('FP8 @50').should('exist');
-      cy.contains('MI355X').should('exist');
-      cy.contains('760').should('exist');
-      cy.contains('-16%').should('exist');
+      cy.get('[data-testid="overview-baseline"][data-hardware="b200"]')
+        .should('contain.text', '432')
+        .and('contain.text', 'FP8');
+      pair('mi355x-vs-b200').within(() => {
+        cy.get('[data-testid="overview-pair-value"][data-hardware="mi355x"]').should(
+          'contain.text',
+          '635',
+        );
+        cy.get('[data-testid="overview-pair-delta"]').should('have.text', '+47% vs B200');
+        cy.contains('At 100, MI355X leads').should('not.exist');
+      });
+      pair('b300-vs-b200').within(() => {
+        cy.get('[data-testid="overview-pair-missing"][data-hardware="b300"]')
+          .should('contain.text', '∞')
+          .and('have.attr', 'title', 'cannot reach @100');
+      });
     });
-    // MiniMax: wider exact-@50 FP8 coverage flips the primary precision to FP8;
-    // FP4 falls back to a single-hardware coverage line.
-    desktopModel('MiniMax-M3').within(() => {
-      cy.contains('FP8 · PRIMARY @50').should('exist');
-      cy.contains('FP4 coverage: H200 measured; insufficient comparable results.').should('exist');
+
+    cy.visit('/overview?tier=30');
+    desktopModel('DeepSeek-V4-Pro').within(() => {
+      pair('b300-vs-b200').within(() => {
+        cy.get('[data-testid="overview-pair-value"][data-hardware="b300"]').should(
+          'contain.text',
+          '1,300',
+        );
+        cy.get('[data-testid="overview-pair-delta"]').should('not.exist');
+      });
+      cy.get('[data-testid="overview-baseline"] [data-testid="overview-pair-missing"]')
+        .should('contain.text', '∞')
+        .and('have.attr', 'title', 'no exact @30 result');
     });
+    cy.get('body')
+      .invoke('text')
+      .should('not.match', /∞\s*%/);
   });
 
-  it('stacks on a 390px phone with no sideways scroll', () => {
-    cy.viewport(390, 844);
-    cy.visit('/overview');
+  it('uses the same cell semantics on mobile and fits both 390px and 320px widths', () => {
+    for (const width of [390, 320]) {
+      cy.viewport(width, 844);
+      cy.visit('/overview');
 
-    cy.get('[data-testid="overview-mobile-list"]').should('be.visible');
-    cy.get('[data-testid="overview-desktop-matrix"]').should('not.be.visible');
-    cy.get('[data-testid="overview-mobile-list"]').within(() => {
-      cy.get('details, summary, button').should('not.exist');
-    });
-    expectNoHorizontalOverflow();
-    expectNoHorizontalScroller('overview-mobile-list');
+      cy.get('[data-testid="overview-mobile-list"]').should('be.visible');
+      cy.get('[data-testid="overview-tier-switcher"]').should('be.visible');
+      cy.get('[data-testid="overview-desktop-matrix"]').should('not.be.visible');
+      mobileModel('Qwen-3.5-397B-A17B').within(() => {
+        cy.get('[data-testid="overview-pair"]').should('have.length', 4);
+        pair('mi355x-vs-b200').within(() => {
+          cy.get('[data-testid="overview-pair-value"][data-hardware="mi355x"]').should(
+            'contain.text',
+            '760',
+          );
+          cy.get('[data-testid="overview-pair-delta"]').should('have.text', '−16% vs B200');
+        });
+      });
+      expectNoHorizontalOverflow();
+      expectNoHorizontalScroller('overview-mobile-list');
+    }
   });
 
-  it('renders the Chinese sibling with the same hierarchy and the per-GPU unit', () => {
-    cy.viewport(1280, 800);
+  it('renders the Chinese sibling with equivalent matrix copy and semantics', () => {
+    cy.viewport(1280, 900);
     cy.visit('/zh/overview');
 
     cy.contains('h1', 'AI 推理总览').should('exist');
-    cy.contains('输出 tok/s/GPU').should('exist');
-    desktopModel('DeepSeek-V4-Pro').within(() => {
-      cy.contains('FP4 · 主排名 @50').should('exist');
+    cy.contains('一眼对比各活跃模型在 MI355X、B200、B300、GB200 与 GB300 上的表现。').should(
+      'exist',
+    );
+    cy.contains('各平台最佳验证配置').should('exist');
+    desktopModel('Qwen-3.5-397B-A17B').within(() => {
+      pair('mi355x-vs-b200').within(() => {
+        cy.get('[data-testid="overview-pair-delta"]').should('have.text', '相对 B200 −16%');
+        cy.contains('100 档由 MI355X 领先').should('exist');
+      });
+      pair('b300-vs-b200').within(() => {
+        cy.get('[data-testid="overview-pair-mismatch"]').should(
+          'have.text',
+          'FP4 与 FP8 · 无可比差值',
+        );
+      });
+    });
+    cy.contains('∞ = 无可比结果').should('exist');
+
+    cy.visit('/zh/overview?tier=100');
+    cy.contains('每 GPU 输出 tok/s @100 tok/s/用户').should('exist');
+    desktopModel('Qwen-3.5-397B-A17B').within(() => {
+      pair('mi355x-vs-b200').within(() => {
+        cy.get('[data-testid="overview-pair-delta"]').should('have.text', '相对 B200 +47%');
+        cy.contains('100 档由 MI355X 领先').should('not.exist');
+      });
+    });
+    cy.get('[data-testid="overview-tier-switcher"]').within(() => {
+      cy.contains('a', '50').should('have.attr', 'href', '/zh/overview');
     });
-    cy.contains('默认精度优先覆盖更多可比较硬件。未参与排名不代表性能更低。').should('exist');
   });
 });
diff --git a/packages/app/cypress/fixtures/api/overview-rows.json b/packages/app/cypress/fixtures/api/overview-rows.json
index d0e53d8a..96a441b0 100644
--- a/packages/app/cypress/fixtures/api/overview-rows.json
+++ b/packages/app/cypress/fixtures/api/overview-rows.json
@@ -1,26 +1,29 @@
 {
   "DeepSeek-V4-Pro": [
-    {"id":1,"hardware":"b300","framework":"sglang","model":"dsv4","precision":"fp4","spec_method":"mtp","disagg":false,"is_multinode":false,"prefill_tp":8,"prefill_ep":1,"prefill_dp_attention":false,"prefill_num_workers":1,"decode_tp":8,"decode_ep":1,"decode_dp_attention":false,"decode_num_workers":1,"num_prefill_gpu":8,"num_decode_gpu":8,"benchmark_type":"single_turn","isl":8192,"osl":1024,"conc":8,"offload_mode":"off","image":null,"metrics":{"median_intvty":30,"output_tput_per_gpu":1300},"date":"2026-06-24","run_url":null},
-    {"id":2,"hardware":"b300","framework":"sglang","model":"dsv4","precision":"fp4","spec_method":"mtp","disagg":false,"is_multinode":false,"prefill_tp":8,"prefill_ep":1,"prefill_dp_attention":false,"prefill_num_workers":1,"decode_tp":8,"decode_ep":1,"decode_dp_attention":false,"decode_num_workers":1,"num_prefill_gpu":8,"num_decode_gpu":8,"benchmark_type":"single_turn","isl":8192,"osl":1024,"conc":16,"offload_mode":"off","image":null,"metrics":{"median_intvty":70,"output_tput_per_gpu":1000},"date":"2026-07-04","run_url":null},
-    {"id":3,"hardware":"b200","framework":"sglang","model":"dsv4","precision":"fp4","spec_method":"mtp","disagg":false,"is_multinode":false,"prefill_tp":8,"prefill_ep":1,"prefill_dp_attention":false,"prefill_num_workers":1,"decode_tp":8,"decode_ep":1,"decode_dp_attention":false,"decode_num_workers":1,"num_prefill_gpu":8,"num_decode_gpu":8,"benchmark_type":"single_turn","isl":8192,"osl":1024,"conc":8,"offload_mode":"off","image":null,"metrics":{"median_intvty":40,"output_tput_per_gpu":1050},"date":"2026-07-18","run_url":null},
-    {"id":4,"hardware":"b200","framework":"sglang","model":"dsv4","precision":"fp4","spec_method":"mtp","disagg":false,"is_multinode":false,"prefill_tp":8,"prefill_ep":1,"prefill_dp_attention":false,"prefill_num_workers":1,"decode_tp":8,"decode_ep":1,"decode_dp_attention":false,"decode_num_workers":1,"num_prefill_gpu":8,"num_decode_gpu":8,"benchmark_type":"single_turn","isl":8192,"osl":1024,"conc":24,"offload_mode":"off","image":null,"metrics":{"median_intvty":110,"output_tput_per_gpu":300},"date":"2026-07-18","run_url":null},
-    {"id":5,"hardware":"mi355x","framework":"sglang","model":"dsv4","precision":"fp4","spec_method":"mtp","disagg":false,"is_multinode":false,"prefill_tp":8,"prefill_ep":1,"prefill_dp_attention":false,"prefill_num_workers":1,"decode_tp":8,"decode_ep":1,"decode_dp_attention":false,"decode_num_workers":1,"num_prefill_gpu":8,"num_decode_gpu":8,"benchmark_type":"single_turn","isl":8192,"osl":1024,"conc":16,"offload_mode":"off","image":null,"metrics":{"median_intvty":60,"output_tput_per_gpu":760},"date":"2026-07-18","run_url":null},
-    {"id":6,"hardware":"gb300","framework":"sglang","model":"dsv4","precision":"fp4","spec_method":"mtp","disagg":false,"is_multinode":false,"prefill_tp":8,"prefill_ep":1,"prefill_dp_attention":false,"prefill_num_workers":1,"decode_tp":8,"decode_ep":1,"decode_dp_attention":false,"decode_num_workers":1,"num_prefill_gpu":8,"num_decode_gpu":8,"benchmark_type":"single_turn","isl":8192,"osl":1024,"conc":8,"offload_mode":"off","image":null,"metrics":{"median_intvty":45,"output_tput_per_gpu":1050},"date":"2026-07-18","run_url":null},
-    {"id":7,"hardware":"h100","framework":"sglang","model":"dsv4","precision":"fp4","spec_method":"none","disagg":false,"is_multinode":false,"prefill_tp":8,"prefill_ep":1,"prefill_dp_attention":false,"prefill_num_workers":1,"decode_tp":8,"decode_ep":1,"decode_dp_attention":false,"decode_num_workers":1,"num_prefill_gpu":8,"num_decode_gpu":8,"benchmark_type":"single_turn","isl":8192,"osl":1024,"conc":8,"offload_mode":"off","image":null,"metrics":{"median_intvty":30,"output_tput_per_gpu":800},"date":"2026-07-18","run_url":null},
-    {"id":8,"hardware":"gb200","framework":"sglang","model":"dsv4","precision":"fp8","spec_method":"mtp","disagg":false,"is_multinode":false,"prefill_tp":8,"prefill_ep":1,"prefill_dp_attention":false,"prefill_num_workers":1,"decode_tp":8,"decode_ep":1,"decode_dp_attention":false,"decode_num_workers":1,"num_prefill_gpu":8,"num_decode_gpu":8,"benchmark_type":"single_turn","isl":8192,"osl":1024,"conc":8,"offload_mode":"off","image":null,"metrics":{"median_intvty":50,"output_tput_per_gpu":600},"date":"2026-07-18","run_url":null}
+    {"id": 1, "hardware": "b300", "framework": "sglang", "model": "dsv4", "precision": "fp4", "spec_method": "mtp", "disagg": false, "is_multinode": false, "prefill_tp": 8, "prefill_ep": 1, "prefill_dp_attention": false, "prefill_num_workers": 1, "decode_tp": 8, "decode_ep": 1, "decode_dp_attention": false, "decode_num_workers": 1, "num_prefill_gpu": 8, "num_decode_gpu": 8, "benchmark_type": "single_turn", "isl": 8192, "osl": 1024, "conc": 8, "offload_mode": "off", "image": null, "metrics": {"median_intvty": 30, "output_tput_per_gpu": 1300}, "date": "2026-06-24", "run_url": null},
+    {"id": 2, "hardware": "b300", "framework": "sglang", "model": "dsv4", "precision": "fp4", "spec_method": "mtp", "disagg": false, "is_multinode": false, "prefill_tp": 8, "prefill_ep": 1, "prefill_dp_attention": false, "prefill_num_workers": 1, "decode_tp": 8, "decode_ep": 1, "decode_dp_attention": false, "decode_num_workers": 1, "num_prefill_gpu": 8, "num_decode_gpu": 8, "benchmark_type": "single_turn", "isl": 8192, "osl": 1024, "conc": 16, "offload_mode": "off", "image": null, "metrics": {"median_intvty": 70, "output_tput_per_gpu": 1000}, "date": "2026-07-04", "run_url": null},
+    {"id": 3, "hardware": "b200", "framework": "sglang", "model": "dsv4", "precision": "fp4", "spec_method": "mtp", "disagg": false, "is_multinode": false, "prefill_tp": 8, "prefill_ep": 1, "prefill_dp_attention": false, "prefill_num_workers": 1, "decode_tp": 8, "decode_ep": 1, "decode_dp_attention": false, "decode_num_workers": 1, "num_prefill_gpu": 8, "num_decode_gpu": 8, "benchmark_type": "single_turn", "isl": 8192, "osl": 1024, "conc": 8, "offload_mode": "off", "image": null, "metrics": {"median_intvty": 40, "output_tput_per_gpu": 1050}, "date": "2026-07-18", "run_url": null},
+    {"id": 4, "hardware": "b200", "framework": "sglang", "model": "dsv4", "precision": "fp4", "spec_method": "mtp", "disagg": false, "is_multinode": false, "prefill_tp": 8, "prefill_ep": 1, "prefill_dp_attention": false, "prefill_num_workers": 1, "decode_tp": 8, "decode_ep": 1, "decode_dp_attention": false, "decode_num_workers": 1, "num_prefill_gpu": 8, "num_decode_gpu": 8, "benchmark_type": "single_turn", "isl": 8192, "osl": 1024, "conc": 24, "offload_mode": "off", "image": null, "metrics": {"median_intvty": 110, "output_tput_per_gpu": 300}, "date": "2026-07-18", "run_url": null},
+    {"id": 5, "hardware": "mi355x", "framework": "sglang", "model": "dsv4", "precision": "fp4", "spec_method": "mtp", "disagg": false, "is_multinode": false, "prefill_tp": 8, "prefill_ep": 1, "prefill_dp_attention": false, "prefill_num_workers": 1, "decode_tp": 8, "decode_ep": 1, "decode_dp_attention": false, "decode_num_workers": 1, "num_prefill_gpu": 8, "num_decode_gpu": 8, "benchmark_type": "single_turn", "isl": 8192, "osl": 1024, "conc": 16, "offload_mode": "off", "image": null, "metrics": {"median_intvty": 60, "output_tput_per_gpu": 760}, "date": "2026-07-18", "run_url": null},
+    {"id": 6, "hardware": "gb300", "framework": "sglang", "model": "dsv4", "precision": "fp4", "spec_method": "mtp", "disagg": false, "is_multinode": false, "prefill_tp": 8, "prefill_ep": 1, "prefill_dp_attention": false, "prefill_num_workers": 1, "decode_tp": 8, "decode_ep": 1, "decode_dp_attention": false, "decode_num_workers": 1, "num_prefill_gpu": 8, "num_decode_gpu": 8, "benchmark_type": "single_turn", "isl": 8192, "osl": 1024, "conc": 8, "offload_mode": "off", "image": null, "metrics": {"median_intvty": 45, "output_tput_per_gpu": 1050}, "date": "2026-07-18", "run_url": null},
+    {"id": 7, "hardware": "h100", "framework": "sglang", "model": "dsv4", "precision": "fp4", "spec_method": "none", "disagg": false, "is_multinode": false, "prefill_tp": 8, "prefill_ep": 1, "prefill_dp_attention": false, "prefill_num_workers": 1, "decode_tp": 8, "decode_ep": 1, "decode_dp_attention": false, "decode_num_workers": 1, "num_prefill_gpu": 8, "num_decode_gpu": 8, "benchmark_type": "single_turn", "isl": 8192, "osl": 1024, "conc": 8, "offload_mode": "off", "image": null, "metrics": {"median_intvty": 30, "output_tput_per_gpu": 800}, "date": "2026-07-18", "run_url": null},
+    {"id": 8, "hardware": "gb200", "framework": "sglang", "model": "dsv4", "precision": "fp8", "spec_method": "mtp", "disagg": false, "is_multinode": false, "prefill_tp": 8, "prefill_ep": 1, "prefill_dp_attention": false, "prefill_num_workers": 1, "decode_tp": 8, "decode_ep": 1, "decode_dp_attention": false, "decode_num_workers": 1, "num_prefill_gpu": 8, "num_decode_gpu": 8, "benchmark_type": "single_turn", "isl": 8192, "osl": 1024, "conc": 8, "offload_mode": "off", "image": null, "metrics": {"median_intvty": 50, "output_tput_per_gpu": 600}, "date": "2026-07-18", "run_url": null}
   ],
   "Qwen-3.5-397B-A17B": [
-    {"id":9,"hardware":"b300","framework":"sglang","model":"qwen3.5","precision":"fp4","spec_method":"mtp","disagg":false,"is_multinode":false,"prefill_tp":8,"prefill_ep":1,"prefill_dp_attention":false,"prefill_num_workers":1,"decode_tp":8,"decode_ep":1,"decode_dp_attention":false,"decode_num_workers":1,"num_prefill_gpu":8,"num_decode_gpu":8,"benchmark_type":"single_turn","isl":8192,"osl":1024,"conc":8,"offload_mode":"off","image":null,"metrics":{"median_intvty":30,"output_tput_per_gpu":1400},"date":"2026-07-18","run_url":null},
-    {"id":10,"hardware":"b300","framework":"sglang","model":"qwen3.5","precision":"fp4","spec_method":"mtp","disagg":false,"is_multinode":false,"prefill_tp":8,"prefill_ep":1,"prefill_dp_attention":false,"prefill_num_workers":1,"decode_tp":8,"decode_ep":1,"decode_dp_attention":false,"decode_num_workers":1,"num_prefill_gpu":8,"num_decode_gpu":8,"benchmark_type":"single_turn","isl":8192,"osl":1024,"conc":16,"offload_mode":"off","image":null,"metrics":{"median_intvty":70,"output_tput_per_gpu":980},"date":"2026-07-18","run_url":null},
-    {"id":11,"hardware":"b200","framework":"sglang","model":"qwen3.5","precision":"fp4","spec_method":"mtp","disagg":false,"is_multinode":false,"prefill_tp":8,"prefill_ep":1,"prefill_dp_attention":false,"prefill_num_workers":1,"decode_tp":8,"decode_ep":1,"decode_dp_attention":false,"decode_num_workers":1,"num_prefill_gpu":8,"num_decode_gpu":8,"benchmark_type":"single_turn","isl":8192,"osl":1024,"conc":8,"offload_mode":"off","image":null,"metrics":{"median_intvty":40,"output_tput_per_gpu":1150},"date":"2026-07-18","run_url":null},
-    {"id":12,"hardware":"b200","framework":"sglang","model":"qwen3.5","precision":"fp4","spec_method":"mtp","disagg":false,"is_multinode":false,"prefill_tp":8,"prefill_ep":1,"prefill_dp_attention":false,"prefill_num_workers":1,"decode_tp":8,"decode_ep":1,"decode_dp_attention":false,"decode_num_workers":1,"num_prefill_gpu":8,"num_decode_gpu":8,"benchmark_type":"single_turn","isl":8192,"osl":1024,"conc":24,"offload_mode":"off","image":null,"metrics":{"median_intvty":80,"output_tput_per_gpu":720},"date":"2026-07-18","run_url":null},
-    {"id":13,"hardware":"b200","framework":"sglang","model":"qwen3.5","precision":"fp8","spec_method":"mtp","disagg":false,"is_multinode":false,"prefill_tp":8,"prefill_ep":1,"prefill_dp_attention":false,"prefill_num_workers":1,"decode_tp":8,"decode_ep":1,"decode_dp_attention":false,"decode_num_workers":1,"num_prefill_gpu":8,"num_decode_gpu":8,"benchmark_type":"single_turn","isl":8192,"osl":1024,"conc":12,"offload_mode":"off","image":null,"metrics":{"median_intvty":50,"output_tput_per_gpu":900},"date":"2026-07-18","run_url":null},
-    {"id":14,"hardware":"mi355x","framework":"sglang","model":"qwen3.5","precision":"fp8","spec_method":"mtp","disagg":false,"is_multinode":false,"prefill_tp":8,"prefill_ep":1,"prefill_dp_attention":false,"prefill_num_workers":1,"decode_tp":8,"decode_ep":1,"decode_dp_attention":false,"decode_num_workers":1,"num_prefill_gpu":8,"num_decode_gpu":8,"benchmark_type":"single_turn","isl":8192,"osl":1024,"conc":12,"offload_mode":"off","image":null,"metrics":{"median_intvty":50,"output_tput_per_gpu":760},"date":"2026-07-18","run_url":null}
+    {"id": 9, "hardware": "b300", "framework": "sglang", "model": "qwen3.5", "precision": "fp4", "spec_method": "mtp", "disagg": false, "is_multinode": false, "prefill_tp": 8, "prefill_ep": 1, "prefill_dp_attention": false, "prefill_num_workers": 1, "decode_tp": 8, "decode_ep": 1, "decode_dp_attention": false, "decode_num_workers": 1, "num_prefill_gpu": 8, "num_decode_gpu": 8, "benchmark_type": "single_turn", "isl": 8192, "osl": 1024, "conc": 8, "offload_mode": "off", "image": null, "metrics": {"median_intvty": 30, "output_tput_per_gpu": 1400}, "date": "2026-07-18", "run_url": null},
+    {"id": 10, "hardware": "b300", "framework": "sglang", "model": "qwen3.5", "precision": "fp4", "spec_method": "mtp", "disagg": false, "is_multinode": false, "prefill_tp": 8, "prefill_ep": 1, "prefill_dp_attention": false, "prefill_num_workers": 1, "decode_tp": 8, "decode_ep": 1, "decode_dp_attention": false, "decode_num_workers": 1, "num_prefill_gpu": 8, "num_decode_gpu": 8, "benchmark_type": "single_turn", "isl": 8192, "osl": 1024, "conc": 16, "offload_mode": "off", "image": null, "metrics": {"median_intvty": 70, "output_tput_per_gpu": 980}, "date": "2026-07-18", "run_url": null},
+    {"id": 11, "hardware": "b200", "framework": "sglang", "model": "qwen3.5", "precision": "fp4", "spec_method": "mtp", "disagg": false, "is_multinode": false, "prefill_tp": 8, "prefill_ep": 1, "prefill_dp_attention": false, "prefill_num_workers": 1, "decode_tp": 8, "decode_ep": 1, "decode_dp_attention": false, "decode_num_workers": 1, "num_prefill_gpu": 8, "num_decode_gpu": 8, "benchmark_type": "single_turn", "isl": 8192, "osl": 1024, "conc": 8, "offload_mode": "off", "image": null, "metrics": {"median_intvty": 40, "output_tput_per_gpu": 800}, "date": "2026-07-18", "run_url": null},
+    {"id": 12, "hardware": "b200", "framework": "sglang", "model": "qwen3.5", "precision": "fp4", "spec_method": "mtp", "disagg": false, "is_multinode": false, "prefill_tp": 8, "prefill_ep": 1, "prefill_dp_attention": false, "prefill_num_workers": 1, "decode_tp": 8, "decode_ep": 1, "decode_dp_attention": false, "decode_num_workers": 1, "num_prefill_gpu": 8, "num_decode_gpu": 8, "benchmark_type": "single_turn", "isl": 8192, "osl": 1024, "conc": 24, "offload_mode": "off", "image": null, "metrics": {"median_intvty": 80, "output_tput_per_gpu": 600}, "date": "2026-07-18", "run_url": null},
+    {"id": 13, "hardware": "b200", "framework": "sglang", "model": "qwen3.5", "precision": "fp8", "spec_method": "mtp", "disagg": false, "is_multinode": false, "prefill_tp": 8, "prefill_ep": 1, "prefill_dp_attention": false, "prefill_num_workers": 1, "decode_tp": 8, "decode_ep": 1, "decode_dp_attention": false, "decode_num_workers": 1, "num_prefill_gpu": 8, "num_decode_gpu": 8, "benchmark_type": "single_turn", "isl": 8192, "osl": 1024, "conc": 12, "offload_mode": "off", "image": null, "metrics": {"median_intvty": 50, "output_tput_per_gpu": 900}, "date": "2026-07-18", "run_url": null},
+    {"id": 14, "hardware": "mi355x", "framework": "sglang", "model": "qwen3.5", "precision": "fp8", "spec_method": "mtp", "disagg": false, "is_multinode": false, "prefill_tp": 8, "prefill_ep": 1, "prefill_dp_attention": false, "prefill_num_workers": 1, "decode_tp": 8, "decode_ep": 1, "decode_dp_attention": false, "decode_num_workers": 1, "num_prefill_gpu": 8, "num_decode_gpu": 8, "benchmark_type": "single_turn", "isl": 8192, "osl": 1024, "conc": 12, "offload_mode": "off", "image": null, "metrics": {"median_intvty": 50, "output_tput_per_gpu": 760}, "date": "2026-07-18", "run_url": null},
+    {"id": 19, "hardware": "b200", "framework": "sglang", "model": "qwen3.5", "precision": "fp8", "spec_method": "mtp", "disagg": false, "is_multinode": false, "prefill_tp": 8, "prefill_ep": 1, "prefill_dp_attention": false, "prefill_num_workers": 1, "decode_tp": 8, "decode_ep": 1, "decode_dp_attention": false, "decode_num_workers": 1, "num_prefill_gpu": 8, "num_decode_gpu": 8, "benchmark_type": "single_turn", "isl": 8192, "osl": 1024, "conc": 24, "offload_mode": "off", "image": null, "metrics": {"median_intvty": 120, "output_tput_per_gpu": 300}, "date": "2026-07-18", "run_url": null},
+    {"id": 20, "hardware": "mi355x", "framework": "sglang", "model": "qwen3.5", "precision": "fp8", "spec_method": "mtp", "disagg": false, "is_multinode": false, "prefill_tp": 8, "prefill_ep": 1, "prefill_dp_attention": false, "prefill_num_workers": 1, "decode_tp": 8, "decode_ep": 1, "decode_dp_attention": false, "decode_num_workers": 1, "num_prefill_gpu": 8, "num_decode_gpu": 8, "benchmark_type": "single_turn", "isl": 8192, "osl": 1024, "conc": 24, "offload_mode": "off", "image": null, "metrics": {"median_intvty": 120, "output_tput_per_gpu": 600}, "date": "2026-07-18", "run_url": null}
   ],
   "MiniMax-M3": [
-    {"id":15,"hardware":"h200","framework":"sglang","model":"minimaxm3","precision":"fp8","spec_method":"mtp","disagg":false,"is_multinode":false,"prefill_tp":8,"prefill_ep":1,"prefill_dp_attention":false,"prefill_num_workers":1,"decode_tp":8,"decode_ep":1,"decode_dp_attention":false,"decode_num_workers":1,"num_prefill_gpu":8,"num_decode_gpu":8,"benchmark_type":"single_turn","isl":8192,"osl":1024,"conc":8,"offload_mode":"off","image":null,"metrics":{"median_intvty":30,"output_tput_per_gpu":820},"date":"2026-07-18","run_url":null},
-    {"id":16,"hardware":"h200","framework":"sglang","model":"minimaxm3","precision":"fp8","spec_method":"mtp","disagg":false,"is_multinode":false,"prefill_tp":8,"prefill_ep":1,"prefill_dp_attention":false,"prefill_num_workers":1,"decode_tp":8,"decode_ep":1,"decode_dp_attention":false,"decode_num_workers":1,"num_prefill_gpu":8,"num_decode_gpu":8,"benchmark_type":"single_turn","isl":8192,"osl":1024,"conc":16,"offload_mode":"off","image":null,"metrics":{"median_intvty":70,"output_tput_per_gpu":560},"date":"2026-07-18","run_url":null},
-    {"id":17,"hardware":"mi325x","framework":"sglang","model":"minimaxm3","precision":"fp8","spec_method":"mtp","disagg":false,"is_multinode":false,"prefill_tp":8,"prefill_ep":1,"prefill_dp_attention":false,"prefill_num_workers":1,"decode_tp":8,"decode_ep":1,"decode_dp_attention":false,"decode_num_workers":1,"num_prefill_gpu":8,"num_decode_gpu":8,"benchmark_type":"single_turn","isl":8192,"osl":1024,"conc":12,"offload_mode":"off","image":null,"metrics":{"median_intvty":50,"output_tput_per_gpu":650},"date":"2026-07-18","run_url":null},
-    {"id":18,"hardware":"h200","framework":"sglang","model":"minimaxm3","precision":"fp4","spec_method":"mtp","disagg":false,"is_multinode":false,"prefill_tp":8,"prefill_ep":1,"prefill_dp_attention":false,"prefill_num_workers":1,"decode_tp":8,"decode_ep":1,"decode_dp_attention":false,"decode_num_workers":1,"num_prefill_gpu":8,"num_decode_gpu":8,"benchmark_type":"single_turn","isl":8192,"osl":1024,"conc":12,"offload_mode":"off","image":null,"metrics":{"median_intvty":50,"output_tput_per_gpu":600},"date":"2026-07-18","run_url":null}
+    {"id": 15, "hardware": "h200", "framework": "sglang", "model": "minimaxm3", "precision": "fp8", "spec_method": "mtp", "disagg": false, "is_multinode": false, "prefill_tp": 8, "prefill_ep": 1, "prefill_dp_attention": false, "prefill_num_workers": 1, "decode_tp": 8, "decode_ep": 1, "decode_dp_attention": false, "decode_num_workers": 1, "num_prefill_gpu": 8, "num_decode_gpu": 8, "benchmark_type": "single_turn", "isl": 8192, "osl": 1024, "conc": 8, "offload_mode": "off", "image": null, "metrics": {"median_intvty": 30, "output_tput_per_gpu": 820}, "date": "2026-07-18", "run_url": null},
+    {"id": 16, "hardware": "h200", "framework": "sglang", "model": "minimaxm3", "precision": "fp8", "spec_method": "mtp", "disagg": false, "is_multinode": false, "prefill_tp": 8, "prefill_ep": 1, "prefill_dp_attention": false, "prefill_num_workers": 1, "decode_tp": 8, "decode_ep": 1, "decode_dp_attention": false, "decode_num_workers": 1, "num_prefill_gpu": 8, "num_decode_gpu": 8, "benchmark_type": "single_turn", "isl": 8192, "osl": 1024, "conc": 16, "offload_mode": "off", "image": null, "metrics": {"median_intvty": 70, "output_tput_per_gpu": 560}, "date": "2026-07-18", "run_url": null},
+    {"id": 17, "hardware": "mi325x", "framework": "sglang", "model": "minimaxm3", "precision": "fp8", "spec_method": "mtp", "disagg": false, "is_multinode": false, "prefill_tp": 8, "prefill_ep": 1, "prefill_dp_attention": false, "prefill_num_workers": 1, "decode_tp": 8, "decode_ep": 1, "decode_dp_attention": false, "decode_num_workers": 1, "num_prefill_gpu": 8, "num_decode_gpu": 8, "benchmark_type": "single_turn", "isl": 8192, "osl": 1024, "conc": 12, "offload_mode": "off", "image": null, "metrics": {"median_intvty": 50, "output_tput_per_gpu": 650}, "date": "2026-07-18", "run_url": null},
+    {"id": 18, "hardware": "h200", "framework": "sglang", "model": "minimaxm3", "precision": "fp4", "spec_method": "mtp", "disagg": false, "is_multinode": false, "prefill_tp": 8, "prefill_ep": 1, "prefill_dp_attention": false, "prefill_num_workers": 1, "decode_tp": 8, "decode_ep": 1, "decode_dp_attention": false, "decode_num_workers": 1, "num_prefill_gpu": 8, "num_decode_gpu": 8, "benchmark_type": "single_turn", "isl": 8192, "osl": 1024, "conc": 12, "offload_mode": "off", "image": null, "metrics": {"median_intvty": 50, "output_tput_per_gpu": 600}, "date": "2026-07-18", "run_url": null},
+    {"id": 21, "hardware": "gb300", "framework": "sglang", "model": "minimaxm3", "precision": "fp8", "spec_method": "mtp", "disagg": false, "is_multinode": false, "prefill_tp": 8, "prefill_ep": 1, "prefill_dp_attention": false, "prefill_num_workers": 1, "decode_tp": 8, "decode_ep": 1, "decode_dp_attention": false, "decode_num_workers": 1, "num_prefill_gpu": 8, "num_decode_gpu": 8, "benchmark_type": "single_turn", "isl": 8192, "osl": 1024, "conc": 12, "offload_mode": "off", "image": null, "metrics": {"median_intvty": 50, "output_tput_per_gpu": 700}, "date": "2026-07-18", "run_url": null}
   ]
 }
diff --git a/packages/app/src/app/overview/page.tsx b/packages/app/src/app/overview/page.tsx
index e76d0d35..8e1bde73 100644
--- a/packages/app/src/app/overview/page.tsx
+++ b/packages/app/src/app/overview/page.tsx
@@ -4,12 +4,13 @@ import { SITE_NAME, SITE_URL } from '@semianalysisai/inferencex-constants';
 
 import { OverviewPageContent } from '@/components/overview/overview-page';
 import { enAlternates } from '@/lib/i18n';
+import { resolveOverviewTier } from '@/lib/overview-data';
 import { getOverviewPageData } from '@/lib/overview-data.server';
 
 export const dynamic = 'force-dynamic';
 
 const DESCRIPTION =
-  'Comparable validated AI inference serving results for every active model and platform at a fixed single-turn 8K input / 1K output workload, ranked at 50 tok/s/user.';
+  'Every active model across MI355X, B200, B300, GB200 and GB300 at a fixed single-turn 8K input / 1K output workload — each platform’s best validated speculative-decode serving result at 50 tok/s/user, with same-precision deltas against B200.';
 
 export const metadata: Metadata = {
   title: 'AI Inference Overview',
@@ -28,8 +29,13 @@ export const metadata: Metadata = {
   },
 };
 
-export default async function OverviewPage() {
-  const data = await getOverviewPageData();
+interface Props {
+  searchParams: Promise<Record<string, string | string[] | undefined>>;
+}
+
+export default async function OverviewPage({ searchParams }: Props) {
+  const sp = await searchParams;
+  const data = await getOverviewPageData(resolveOverviewTier(sp.tier));
   return (
     <main className="relative">
       <div className="container mx-auto px-4 lg:px-8 pb-8">
diff --git a/packages/app/src/app/zh/overview/page.tsx b/packages/app/src/app/zh/overview/page.tsx
index 91e4fd48..dd5d6cc9 100644
--- a/packages/app/src/app/zh/overview/page.tsx
+++ b/packages/app/src/app/zh/overview/page.tsx
@@ -4,12 +4,13 @@ import { SITE_NAME, SITE_URL } from '@semianalysisai/inferencex-constants';
 
 import { OverviewPageContent } from '@/components/overview/overview-page';
 import { ZH_OG_LOCALE, zhAlternates } from '@/lib/i18n';
+import { resolveOverviewTier } from '@/lib/overview-data';
 import { getOverviewPageData } from '@/lib/overview-data.server';
 
 export const dynamic = 'force-dynamic';
 
 const DESCRIPTION =
-  '以固定单轮 8K 输入 / 1K 输出负载，按 50 tok/s/user 档位排名，对比每个活跃模型在各硬件平台上可比的已验证服务结果。';
+  '在固定单轮 8K 输入 / 1K 输出负载下，总览各活跃模型在 MI355X、B200、B300、GB200 与 GB300 上的表现；每格为该平台最佳验证推测解码结果（50 tok/s/user 档位），相对 B200 的差值仅在同精度结果之间计算。';
 
 export const metadata: Metadata = {
   title: 'AI 推理总览',
@@ -29,8 +30,13 @@ export const metadata: Metadata = {
   },
 };
 
-export default async function ZhOverviewPage() {
-  const data = await getOverviewPageData();
+interface Props {
+  searchParams: Promise<Record<string, string | string[] | undefined>>;
+}
+
+export default async function ZhOverviewPage({ searchParams }: Props) {
+  const sp = await searchParams;
+  const data = await getOverviewPageData(resolveOverviewTier(sp.tier));
   return (
     <main className="relative">
       <div className="container mx-auto px-4 lg:px-8 pb-8">
diff --git a/packages/app/src/components/overview/overview-page.tsx b/packages/app/src/components/overview/overview-page.tsx
index c99f3ae6..5ac8c864 100644
--- a/packages/app/src/components/overview/overview-page.tsx
+++ b/packages/app/src/components/overview/overview-page.tsx
@@ -5,6 +5,7 @@ import {
   DesktopOverviewMatrix,
   MobileOverviewList,
   OverviewMethodology,
+  OverviewTierSwitcher,
   overviewFormatters,
   OVERVIEW_STRINGS,
   type OverviewLocale,
@@ -31,11 +32,12 @@ export function OverviewPageContent({ data, locale }: OverviewPageProps) {
           {strings.purpose}
         </p>
         <p className="mt-1.5 max-w-4xl text-xs leading-snug text-muted-foreground sm:text-sm">
-          {strings.scope}
+          {strings.scope(data.tier)}
         </p>
         {snapshot === null ? null : (
           <p className="mt-1 text-xs text-muted-foreground/80 tabular-nums">{snapshot}</p>
         )}
+        <OverviewTierSwitcher tier={data.tier} locale={locale} strings={strings} />
       </header>
 
       {/* Official-only summary; uploaded runs remain in the linked dashboard. */}
diff --git a/packages/app/src/components/overview/overview-scorecard.tsx b/packages/app/src/components/overview/overview-scorecard.tsx
index f80c1d07..b5d09806 100644
--- a/packages/app/src/components/overview/overview-scorecard.tsx
+++ b/packages/app/src/components/overview/overview-scorecard.tsx
@@ -1,149 +1,100 @@
 /**
  * @file overview-scorecard.tsx
- * @description The comparison surface itself. Desktop (a matrix at `xl`) and
- * mobile (stacked cards below `xl`) render the SAME per-model hierarchy from one
- * `ModelHierarchy` component, so the two surfaces cannot drift semantically.
+ * @description The comparison surface itself: a portfolio matrix of every
+ * active model (rows) across today's key serving platforms (columns — B200 as
+ * the reference, then MI355X, B300, GB200 NVL72, GB300 NVL72). The desktop
+ * table at `xl` and the mobile cards below `xl` render every cell through the
+ * SAME `BaselineCell`/`CandidateCell` components, so the two surfaces cannot
+ * drift semantically.
  *
- * Each model reads top to bottom as: the primary precision's ranked hardware as
- * horizontal chips (@50), the secondary precision (a subordinate ranked block or
- * one compact coverage line), the @100 capability read, then a not-ranked line
- * that accounts for every remaining hardware with an explicit reason. Precisions
- * and incompatible cohorts are never compared across — each ranks only itself.
+ * Each cell shows that platform's own best validated speculative read at the
+ * displayed tier (50 tok/s/user by default; `?tier=` picks another) with its
+ * precision badge, evidence date and filtered-dashboard link. The signed
+ * delta against B200 appears only when both displayed reads share a precision
+ * and a release — mismatches say so instead of comparing FP4 with FP8. A side
+ * with no exact read shows `∞` with its reason, never a number or percent.
  */
 
-import { Badge } from '@/components/ui/badge';
 import {
-  overviewHighValue,
   OVERVIEW_HIGH_TIER,
-  OVERVIEW_PRIMARY_TIER,
-  type OverviewComparisonGroup,
-  type OverviewConfigResult,
-  type OverviewHardwareStatus,
+  OVERVIEW_TIERS,
+  type OverviewHeadlinePairComparison,
+  type OverviewHeadlinePairMember,
   type OverviewModelSummary,
-  type OverviewNotRankedEntry,
+  type OverviewTier,
 } from '@/lib/overview-data';
-import { detailHref } from '@/lib/overview-links';
+import { buildOverviewDashboardHref, detailHref, overviewTierHref } from '@/lib/overview-links';
 
-import { OverviewDetailLink, OverviewDashboardLink } from './overview-detail-link';
+import { OverviewDetailLink } from './overview-detail-link';
 
 export type OverviewLocale = 'en' | 'zh';
 
 export const OVERVIEW_STRINGS = {
   en: {
     title: 'AI Inference Overview',
-    purpose: 'Compare validated serving results across active models and hardware.',
-    scope: `8K→1K · Single-turn · Output tok/s/GPU at ${OVERVIEW_PRIMARY_TIER} tok/s/user · Speculative decode only · FP4/FP8 by comparable coverage`,
+    purpose: 'Every active model across MI355X, B200, B300, GB200 and GB300 at a glance.',
+    scope: (tier: number) =>
+      `8K→1K · Single-turn · Output tok/s/GPU @${tier} tok/s/user · Speculative decode only · Best validated stack per platform`,
+    tierNavLabel: 'Service level',
+    tierUnit: 'tok/s/user',
     snapshot: (through: string) => `Database snapshot through ${through}`,
+    caption:
+      "Best validated serving results for every active model across today's key platforms, each compared with B200.",
     modelHeader: 'Model',
-    resultsHeader: 'Ranked results',
     detailsHeader: 'Details',
-    caption: `Validated output tok/s/GPU per model and comparable cohort at ${OVERVIEW_PRIMARY_TIER} and ${OVERVIEW_HIGH_TIER} tok/s/user.`,
-    cohort: (engine: string, mode: string) => `${engine} · ${mode}`,
-    aggregated: 'Aggregated',
-    disaggregated: 'Disaggregated',
-    anyEngine: 'All engines',
-    engineGroupLabels: {
-      trt: 'TRTLLM family',
-      sglang: 'SGLang/ATOM family',
-      vllm: 'vLLM family',
-    } as Record<string, string>,
-    engineGroupFallback: (engineGroup: string) => `${engineGroup.toUpperCase()} family`,
-    primaryHeading: (precision: string) => `${precision} · PRIMARY @${OVERVIEW_PRIMARY_TIER}`,
-    secondaryHeading: (precision: string) => `${precision} @${OVERVIEW_PRIMARY_TIER}`,
-    secondaryCoverage: (precision: string, list: string) =>
-      `${precision} coverage: ${list} measured; insufficient comparable results.`,
-    highHeading: `@${OVERVIEW_HIGH_TIER}`,
-    leader: 'Leader',
-    onlyExactResult: 'Only exact result',
-    highLeader: (hardware: string, value: string) => `${hardware} · ${value}`,
-    highLeaderSame: 'Same leader',
-    leaderChange: 'Leader change',
-    highInsufficient: `No exact read at ${OVERVIEW_HIGH_TIER} tok/s/user`,
-    highNoPrimaryBaseline: `No ${OVERVIEW_PRIMARY_TIER} tok/s/user baseline`,
-    evidenceHighDashboard: `Dashboard @${OVERVIEW_HIGH_TIER}`,
-    evidenceDashboardAria: (model: string, config: string) =>
-      `Open filtered dashboard: ${model} · ${config}`,
-    evidenceHighAria: (model: string, config: string) =>
-      `Open filtered dashboard @${OVERVIEW_HIGH_TIER} tok/s/user leader: ${model} · ${config}`,
+    detailLink: 'View details',
+    detailAria: (modelLabel: string) => `View details: ${modelLabel}`,
+    dashboardAria: (modelLabel: string, stack: string) =>
+      `Open filtered dashboard: ${modelLabel} · ${stack}`,
+    delta: (signedPct: string) => `${signedPct} vs B200`,
+    precisionMismatch: (candidatePrecision: string, baselinePrecision: string) =>
+      `${candidatePrecision.toUpperCase()} vs ${baselinePrecision.toUpperCase()} · no comparable delta`,
+    versionMismatch: 'Different releases · no comparable delta',
+    highLeadChanged: (leaderLabel: string) => `At ${OVERVIEW_HIGH_TIER}, ${leaderLabel} leads`,
     infinityLegend: '∞ = no comparable result',
-    notRankedReasons: {
+    missingReasons: (tier: number): Record<string, string> => ({
       standard_decode_only: 'standard decode only',
       int4_bf16_only: 'INT4/BF16 only',
-      different_serving_cohort: 'different serving cohort',
       no_8k1k_data: 'no 8K/1K data',
-      cannot_reach_at50: `cannot reach @${OVERVIEW_PRIMARY_TIER}`,
-      no_exact_at50: `no exact @${OVERVIEW_PRIMARY_TIER} result`,
-    } as Record<string, string>,
-    otherPrecisionReason: (precisions: string[]) =>
-      `${precisions.map((precision) => precision.toUpperCase()).join('/')} only`,
-    joinList: (labels: string[]) =>
-      labels.length <= 1
-        ? (labels[0] ?? '')
-        : `${labels.slice(0, -1).join(', ')} and ${labels.at(-1) ?? ''}`,
+      cannot_reach_at_tier: `cannot reach @${tier}`,
+      no_exact_at_tier: `no exact @${tier} result`,
+    }),
     methodologyNote:
-      'Precisions and incompatible serving cohorts are ranked separately. The default precision maximizes comparable hardware coverage. Not ranked does not mean slower.',
+      "Each cell shows the platform's best validated speculative-decode serving configuration for that model, labeled with its precision. Deltas against B200 compare only same-precision, same-release results — FP4 is never measured against FP8. Results compare complete serving stacks rather than isolated silicon.",
     interpolationNote:
       'Tier values interpolate each configuration’s official Pareto frontier — no extrapolation.',
-    cohortNote:
-      'Engine families with different speculative-decode acceptance forcing, and aggregated versus disaggregated deployments, rank as separate cohorts — never against each other.',
-    detailLink: 'View details',
-    detailAria: (modelLabel: string) => `View details: ${modelLabel}`,
   },
   zh: {
     title: 'AI 推理总览',
-    purpose: '对比活跃模型在不同硬件上的已验证服务结果。',
-    scope: `8K→1K · 单轮 · ${OVERVIEW_PRIMARY_TIER} tok/s/user 下的输出 tok/s/GPU · 仅投机解码 · 按可比覆盖选择 FP4/FP8`,
+    purpose: '一眼对比各活跃模型在 MI355X、B200、B300、GB200 与 GB300 上的表现。',
+    scope: (tier: number) =>
+      `8K→1K · 单轮 · 每 GPU 输出 tok/s @${tier} tok/s/用户 · 仅推测解码 · 各平台最佳验证配置`,
+    tierNavLabel: '服务档位',
+    tierUnit: 'tok/s/用户',
     snapshot: (through: string) => `数据库快照截至 ${through}`,
+    caption: '各活跃模型在关键平台上的最佳验证服务结果，均与 B200 对比。',
     modelHeader: '模型',
-    resultsHeader: '排名结果',
     detailsHeader: '详情',
-    caption: `各模型与可对比分组在 ${OVERVIEW_PRIMARY_TIER} 与 ${OVERVIEW_HIGH_TIER} tok/s/user 下的已验证输出 tok/s/GPU。`,
-    cohort: (engine: string, mode: string) => `${engine} · ${mode}`,
-    aggregated: '聚合部署',
-    disaggregated: '分离部署',
-    anyEngine: '全部引擎',
-    engineGroupLabels: {
-      trt: 'TRTLLM 系列',
-      sglang: 'SGLang/ATOM 系列',
-      vllm: 'vLLM 系列',
-    } as Record<string, string>,
-    engineGroupFallback: (engineGroup: string) => `${engineGroup.toUpperCase()} 系列`,
-    primaryHeading: (precision: string) => `${precision} · 主排名 @${OVERVIEW_PRIMARY_TIER}`,
-    secondaryHeading: (precision: string) => `${precision} @${OVERVIEW_PRIMARY_TIER}`,
-    secondaryCoverage: (precision: string, list: string) =>
-      `${precision} 覆盖：已测量 ${list}；可比结果不足。`,
-    highHeading: `@${OVERVIEW_HIGH_TIER}`,
-    leader: '领先',
-    onlyExactResult: '唯一精确读数',
-    highLeader: (hardware: string, value: string) => `${hardware} · ${value}`,
-    highLeaderSame: '领先者相同',
-    leaderChange: '领先者变化',
-    highInsufficient: `${OVERVIEW_HIGH_TIER} tok/s/user 下无精确读数`,
-    highNoPrimaryBaseline: `无 ${OVERVIEW_PRIMARY_TIER} tok/s/user 基线`,
-    evidenceHighDashboard: `仪表板 @${OVERVIEW_HIGH_TIER}`,
-    evidenceDashboardAria: (model: string, config: string) =>
-      `打开筛选后的仪表板：${model} · ${config}`,
-    evidenceHighAria: (model: string, config: string) =>
-      `打开筛选后的仪表板 @${OVERVIEW_HIGH_TIER} tok/s/user：${model} · ${config}`,
+    detailLink: '查看详情',
+    detailAria: (modelLabel: string) => `查看详情：${modelLabel}`,
+    dashboardAria: (modelLabel: string, stack: string) =>
+      `打开筛选后的仪表板：${modelLabel} · ${stack}`,
+    delta: (signedPct: string) => `相对 B200 ${signedPct}`,
+    precisionMismatch: (candidatePrecision: string, baselinePrecision: string) =>
+      `${candidatePrecision.toUpperCase()} 与 ${baselinePrecision.toUpperCase()} · 无可比差值`,
+    versionMismatch: '版本不同 · 无可比差值',
+    highLeadChanged: (leaderLabel: string) => `${OVERVIEW_HIGH_TIER} 档由 ${leaderLabel} 领先`,
     infinityLegend: '∞ = 无可比结果',
-    notRankedReasons: {
+    missingReasons: (tier: number): Record<string, string> => ({
       standard_decode_only: '仅标准解码',
       int4_bf16_only: '仅 INT4/BF16',
-      different_serving_cohort: '不同服务配置',
       no_8k1k_data: '无 8K/1K 数据',
-      cannot_reach_at50: `无法达到 @${OVERVIEW_PRIMARY_TIER}`,
-      no_exact_at50: `无精确 @${OVERVIEW_PRIMARY_TIER} 结果`,
-    } as Record<string, string>,
-    otherPrecisionReason: (precisions: string[]) =>
-      `仅 ${precisions.map((precision) => precision.toUpperCase()).join('/')}`,
-    joinList: (labels: string[]) => labels.join('、'),
+      cannot_reach_at_tier: `无法达到 @${tier}`,
+      no_exact_at_tier: `无精确 @${tier} 结果`,
+    }),
     methodologyNote:
-      '不同精度和不可直接比较的服务配置将分别排名。默认精度优先覆盖更多可比较硬件。未参与排名不代表性能更低。',
+      '每格展示该平台在该模型上表现最好的已验证推测解码服务配置，并标注精度。相对 B200 的差值只在同精度、同版本结果之间计算——绝不将 FP4 与 FP8 互比。对比对象是完整服务栈，而非单独的芯片。',
     interpolationNote: '各档位数据基于各配置官方 Pareto 前沿插值；不进行外推。',
-    cohortNote:
-      '投机解码接受率强制方式不同的引擎系列，以及聚合与分离部署，按独立分组排名，彼此之间不作比较。',
-    detailLink: '查看详情',
-    detailAria: (modelLabel: string) => `查看详情：${modelLabel}`,
   },
 } as const;
 
@@ -151,19 +102,11 @@ export type OverviewStrings = (typeof OVERVIEW_STRINGS)[OverviewLocale];
 
 interface Formatters {
   number: Intl.NumberFormat;
-  signed: Intl.NumberFormat;
-  date: (date: string) => string;
   shortDate: (date: string) => string;
 }
 
 export function overviewFormatters(locale: OverviewLocale): Formatters {
   const tag = locale === 'zh' ? 'zh-CN' : 'en-US';
-  const dateFormat = new Intl.DateTimeFormat(tag, {
-    year: 'numeric',
-    month: 'short',
-    day: 'numeric',
-    timeZone: 'UTC',
-  });
   const shortDateFormat = new Intl.DateTimeFormat(tag, {
     month: 'short',
     day: 'numeric',
@@ -171,8 +114,6 @@ export function overviewFormatters(locale: OverviewLocale): Formatters {
   });
   return {
     number: new Intl.NumberFormat(tag, { maximumFractionDigits: 0 }),
-    signed: new Intl.NumberFormat(tag, { maximumFractionDigits: 0, signDisplay: 'exceptZero' }),
-    date: (date) => dateFormat.format(new Date(`${date}T00:00:00Z`)),
     shortDate: (date) => shortDateFormat.format(new Date(`${date}T00:00:00Z`)),
   };
 }
@@ -193,448 +134,192 @@ function formatEvidenceDate(
     : `${from}–${formatters.shortDate(evidenceDate.to)}`;
 }
 
-function cohortLabel(
-  group: OverviewComparisonGroup,
-  strings: OverviewStrings,
-  showDbModel: boolean,
-): string {
-  const engine =
-    group.engineGroup === null
-      ? strings.anyEngine
-      : (strings.engineGroupLabels[group.engineGroup] ??
-        strings.engineGroupFallback(group.engineGroup));
-  const mode =
-    group.deploymentMode === 'disaggregated' ? strings.disaggregated : strings.aggregated;
-  const label = strings.cohort(engine, mode);
-  // Point releases of one display model rank in separate cohorts; name the raw
-  // db model so two same-engine, same-mode cohorts are told apart.
-  return showDbModel ? `${label} · ${group.dbModel}` : label;
-}
-
-/** Cohorts with at least one exact @50 read — the ones that render ranked chips. */
-function rankedCohorts(groups: readonly OverviewComparisonGroup[]): OverviewComparisonGroup[] {
-  return groups.filter((group) => group.primaryRanking.state !== 'insufficient_coverage');
+/** The reason copy for a missing side, keyed to the displayed tier. */
+function missingReasonCopy(member: OverviewHeadlinePairMember, strings: OverviewStrings): string {
+  const reason = member.missingReason;
+  return reason === null ? '' : strings.missingReasons(member.read.tier)[reason];
 }
 
-/** A cohort's hardware that hold an exact @50 read, leader first then value desc. */
-function rankedStatuses(group: OverviewComparisonGroup): OverviewHardwareStatus[] {
-  return group.hardwareStatuses
-    .filter((status) => status.isPrimaryLeader || status.primaryDeltaPercent !== null)
-    .toSorted((a, b) => (b.primary.value ?? -1) - (a.primary.value ?? -1));
-}
-
-/** Not-ranked reason copy; `other_precision_only` names the measured precisions. */
-function notRankedReason(entry: OverviewNotRankedEntry, strings: OverviewStrings): string {
-  return entry.reason === 'other_precision_only'
-    ? strings.otherPrecisionReason(entry.precisions)
-    : strings.notRankedReasons[entry.reason];
-}
-
-/** The exact deployment topology of a ranked configuration, for an aria-label. */
-function evidenceConfigLabel(
-  config: OverviewConfigResult,
-  strings: OverviewStrings,
-  showDbModel: boolean,
-): string {
-  return [
-    config.hardwareLabel,
-    config.precision.toUpperCase(),
-    config.frameworkLabel,
-    config.specLabel,
-    config.disagg ? strings.disaggregated : strings.aggregated,
-    config.parallelism,
-    `${config.totalGpu} GPU`,
-    ...(showDbModel ? [config.dbModel] : []),
-  ].join(' · ');
+/** `candidate/B200 − 1` as a rounded percent with a U+2212 or `+` sign. */
+function signedPercent(formatters: Formatters, percent: number): string {
+  const rounded = Math.round(percent);
+  const sign = rounded < 0 ? '−' : '+';
+  return `${sign}${formatters.number.format(Math.abs(rounded))}%`;
 }
 
-const BLOCK_HEADING_CLASS =
-  'mb-0.5 text-xs font-semibold uppercase tracking-wider text-muted-foreground';
-const SUBORDINATE_HEADING_CLASS =
-  'mb-0.5 text-[11px] font-medium uppercase tracking-wider text-muted-foreground/80';
-
 /**
- * One ranked hardware: its label, its @50 value linking to the filtered
- * dashboard, and its standing — `Leader`, its signed delta against the leader,
- * or `Only exact result` when it is the cohort's lone exact read.
+ * The @100 line, rendered only when the @100 leader is a different hardware
+ * than the displayed tier's leader — the leader being the side with the
+ * greater exact @100 read. Null on the 100 view itself (data layer).
  */
-function RankedChip({
-  locale,
-  model,
-  group,
-  status,
-  multiDbModel,
-  formatters,
-  strings,
-}: {
-  locale: OverviewLocale;
-  model: OverviewModelSummary;
-  group: OverviewComparisonGroup;
-  status: OverviewHardwareStatus;
-  multiDbModel: boolean;
-  formatters: Formatters;
-  strings: OverviewStrings;
-}) {
-  const { value, config } = status.primary;
-  const standing = status.isPrimaryLeader
-    ? group.primaryRanking.state === 'single_measured'
-      ? strings.onlyExactResult
-      : strings.leader
-    : `${formatters.signed.format(status.primaryDeltaPercent as number)}%`;
-  const date = formatEvidenceDate(formatters, status.primary.evidenceDate);
-  return (
-    <p className="flex flex-wrap items-baseline gap-x-1.5 gap-y-0.5">
-      <span className="font-medium">{status.hardwareLabel}</span>
-      {value === null ? null : config === null ? (
-        <span className="text-sm font-semibold tabular-nums">
-          {formatters.number.format(value)}
-        </span>
-      ) : (
-        <OverviewDashboardLink
-          locale={locale}
-          model={model}
-          config={config}
-          ariaLabel={strings.evidenceDashboardAria(
-            model.modelLabel,
-            evidenceConfigLabel(config, strings, multiDbModel),
-          )}
-        >
-          <span className="text-sm font-semibold tabular-nums text-foreground">
-            {formatters.number.format(value)}
-          </span>
-        </OverviewDashboardLink>
-      )}
-      <span className="text-xs text-muted-foreground tabular-nums">· {standing}</span>
-      {date === null ? null : (
-        <span className="text-xs text-muted-foreground/80 tabular-nums">· {date}</span>
-      )}
-    </p>
-  );
+function highLeadLine(
+  pair: OverviewHeadlinePairComparison,
+  strings: OverviewStrings,
+): string | null {
+  if (pair.highLeaderTransition !== 'changed_hardware') return null;
+  const candidateHigh = pair.candidate.highRead.value;
+  const baselineHigh = pair.baseline.highRead.value;
+  if (candidateHigh === null || baselineHigh === null) return null;
+  const leaderLabel =
+    candidateHigh > baselineHigh ? pair.candidate.hardwareLabel : pair.baseline.hardwareLabel;
+  return strings.highLeadChanged(leaderLabel);
 }
 
-/** A precision's ranked cohorts as stacked chips, under one heading. */
-function RankedSection({
-  locale,
-  model,
-  cohorts,
-  heading,
-  headingClass,
-  formatters,
-  strings,
-}: {
-  locale: OverviewLocale;
-  model: OverviewModelSummary;
-  cohorts: OverviewComparisonGroup[];
-  heading: string;
-  headingClass: string;
-  formatters: Formatters;
-  strings: OverviewStrings;
-}) {
-  if (cohorts.length === 0) return null;
-  const multiDbModel = new Set(cohorts.map((group) => group.dbModel)).size > 1;
-  return (
-    <div>
-      <p className={headingClass}>{heading}</p>
-      {cohorts.map((group) => (
-        <div key={group.id} className="space-y-0.5">
-          {cohorts.length > 1 ? (
-            <p className="text-xs text-muted-foreground">
-              {cohortLabel(group, strings, multiDbModel)}
-            </p>
-          ) : null}
-          {rankedStatuses(group).map((status) => (
-            <RankedChip
-              key={status.hardware}
-              locale={locale}
-              model={model}
-              group={group}
-              status={status}
-              multiDbModel={multiDbModel}
-              formatters={formatters}
-              strings={strings}
-            />
-          ))}
-        </div>
-      ))}
-    </div>
-  );
-}
+const PAIR_VALUE_LINK_CLASS =
+  'inline-flex min-h-11 items-center rounded-sm underline decoration-dotted underline-offset-4 hover:decoration-solid focus-visible:outline-none focus-visible:ring-[3px] focus-visible:ring-ring/50';
 
-/** The primary precision block: its heading names the precision and the tier. */
-function PrimarySection({
-  locale,
-  model,
-  cohorts,
-  formatters,
-  strings,
-}: {
-  locale: OverviewLocale;
-  model: OverviewModelSummary;
-  cohorts: OverviewComparisonGroup[];
-  formatters: Formatters;
-  strings: OverviewStrings;
-}) {
-  const precision = (model.selectedPrecision ?? '').toUpperCase();
+/** A side with no exact read at the displayed tier: `∞` plus its reason, no link/date/percent. */
+function CellMissing({ hardware, reason }: { hardware: string; reason: string }) {
   return (
-    <RankedSection
-      locale={locale}
-      model={model}
-      cohorts={cohorts}
-      heading={strings.primaryHeading(precision)}
-      headingClass={BLOCK_HEADING_CLASS}
-      formatters={formatters}
-      strings={strings}
-    />
+    <span
+      data-testid="overview-pair-missing"
+      data-hardware={hardware}
+      title={reason}
+      className="inline-flex items-baseline gap-1 text-muted-foreground"
+    >
+      <span aria-hidden="true">{'∞'}</span>
+      <span className="sr-only">{reason}</span>
+    </span>
   );
 }
 
 /**
- * The secondary precision, subordinate to the primary: a ranked block when it
- * genuinely adds comparable hardware, otherwise one compact coverage line so the
- * precision still surfaces without a second full table.
+ * One platform's read: its exact displayed-tier value linked to the platform's
+ * own filtered dashboard, with a precision badge and the evidence date — or
+ * `∞` with the reason it has no exact read.
  */
-function SecondarySection({
+function CellValue({
   locale,
   model,
+  member,
   formatters,
   strings,
 }: {
   locale: OverviewLocale;
   model: OverviewModelSummary;
+  member: OverviewHeadlinePairMember;
   formatters: Formatters;
   strings: OverviewStrings;
 }) {
-  const secondary = model.secondary;
-  if (secondary === null) return null;
-  const precision = secondary.precision.toUpperCase();
-  if (secondary.state === 'coverage') {
-    const list = strings.joinList(secondary.measuredHardware);
-    if (list === '') return null;
-    return (
-      <p className="text-xs text-muted-foreground">{strings.secondaryCoverage(precision, list)}</p>
-    );
+  const { value, config, evidenceDate } = member.read;
+  if (member.missingReason !== null || value === null) {
+    return <CellMissing hardware={member.hardware} reason={missingReasonCopy(member, strings)} />;
   }
+  // The full deployment stack behind this read; the framework and precision
+  // stay visible, the spec method rides the link title and aria only.
+  const stack =
+    config === null
+      ? null
+      : `${member.hardwareLabel} · ${config.frameworkLabel} · ${config.precision.toUpperCase()} · ${config.specLabel}`;
   return (
-    <RankedSection
-      locale={locale}
-      model={model}
-      cohorts={rankedCohorts(secondary.comparisonGroups)}
-      heading={strings.secondaryHeading(precision)}
-      headingClass={SUBORDINATE_HEADING_CLASS}
-      formatters={formatters}
-      strings={strings}
-    />
-  );
-}
-
-/**
- * The 100 tok/s/user read of one cohort. It is ranked from its own evidence, so
- * it names its own leader and says plainly whether that leader changed.
- */
-function HighTierRead({
-  group,
-  formatters,
-  strings,
-}: {
-  group: OverviewComparisonGroup;
-  formatters: Formatters;
-  strings: OverviewStrings;
-}) {
-  const leader = group.highRanking.leader;
-  const value = leader === null ? null : overviewHighValue(leader);
-  if (leader === null || value === null) {
-    return <span className="text-xs text-muted-foreground">{strings.highInsufficient}</span>;
-  }
-  // The @100 leader's own read date — from its hardware's high read, so a @100
-  // read backed by a different config than the @50 leader shows its own date.
-  const date = formatEvidenceDate(
-    formatters,
-    group.hardwareStatuses.find((status) => status.hardware === leader.hardware)?.high
-      .evidenceDate ?? null,
-  );
-  return (
-    <span className="flex flex-wrap items-baseline gap-x-2 gap-y-0.5">
-      <span className="tabular-nums">
-        {strings.highLeader(leader.hardwareLabel, formatters.number.format(value))}
+    <div className="flex flex-wrap items-center gap-x-2 gap-y-0.5 text-sm">
+      <span
+        data-testid="overview-pair-value"
+        data-hardware={member.hardware}
+        className="font-semibold tabular-nums"
+      >
+        {config === null || stack === null ? (
+          formatters.number.format(value)
+        ) : (
+          <a
+            href={buildOverviewDashboardHref(locale, model, config)}
+            title={stack}
+            aria-label={strings.dashboardAria(model.modelLabel, stack)}
+            className={PAIR_VALUE_LINK_CLASS}
+          >
+            {formatters.number.format(value)}
+          </a>
+        )}
       </span>
-      {group.highRanking.state === 'single_measured' ? (
-        <span className="text-xs text-muted-foreground">{strings.onlyExactResult}</span>
-      ) : group.highLeaderTransition === 'changed_hardware' ? (
-        <Badge
-          variant="outline"
-          className="border-amber-500/40 bg-amber-500/10 text-amber-700 dark:border-amber-400/30 dark:text-amber-400"
-          title={strings.leaderChange}
-        >
-          {strings.leaderChange}
-        </Badge>
-      ) : group.highLeaderTransition === 'no_primary_baseline' ? (
-        <span className="text-xs text-muted-foreground">{strings.highNoPrimaryBaseline}</span>
-      ) : (
-        <span className="text-xs text-muted-foreground">{strings.highLeaderSame}</span>
+      {member.precision === null ? null : (
+        <span className="text-[10px] font-semibold uppercase tracking-wider text-muted-foreground">
+          {config === null
+            ? member.precision.toUpperCase()
+            : `${config.frameworkLabel} · ${config.precision.toUpperCase()}`}
+        </span>
       )}
-      {date === null ? null : (
-        <span className="text-xs text-muted-foreground/80 tabular-nums">· {date}</span>
+      {evidenceDate === null ? null : (
+        <span
+          data-testid="overview-pair-evidence-date"
+          data-hardware={member.hardware}
+          className="text-xs text-muted-foreground/80 tabular-nums"
+        >
+          {formatEvidenceDate(formatters, evidenceDate)}
+        </span>
       )}
-    </span>
+    </div>
   );
 }
 
-/**
- * The @100 capability read per comparable cohort. When the @100 leader runs a
- * different configuration than the @50 leader, its own filtered dashboard rides
- * along so the high-tier read ships with the evidence behind it.
- */
-function HighSection({
-  locale,
-  model,
-  cohorts,
-  formatters,
-  strings,
-}: {
+/** The B200 reference cell: the read alone, no delta against itself. */
+function BaselineCell(props: {
   locale: OverviewLocale;
   model: OverviewModelSummary;
-  cohorts: OverviewComparisonGroup[];
+  member: OverviewHeadlinePairMember;
   formatters: Formatters;
   strings: OverviewStrings;
 }) {
-  const highCohorts = cohorts.filter((group) => group.primaryRanking.state === 'comparable');
-  if (highCohorts.length === 0) return null;
-  const multiDbModel = new Set(highCohorts.map((group) => group.dbModel)).size > 1;
   return (
-    <div>
-      <p className={BLOCK_HEADING_CLASS}>{strings.highHeading}</p>
-      {highCohorts.map((group) => {
-        const primaryLeader = group.primaryRanking.leader;
-        const highLeader = group.highRanking.leader;
-        return (
-          <div key={group.id} className="flex flex-wrap items-baseline gap-x-3 gap-y-0.5 text-sm">
-            {highCohorts.length > 1 ? (
-              <span className="text-xs text-muted-foreground">
-                {cohortLabel(group, strings, multiDbModel)}
-              </span>
-            ) : null}
-            <HighTierRead group={group} formatters={formatters} strings={strings} />
-            {highLeader !== null &&
-            (primaryLeader === null || highLeader.key !== primaryLeader.key) ? (
-              <OverviewDashboardLink
-                locale={locale}
-                model={model}
-                config={highLeader}
-                ariaLabel={strings.evidenceHighAria(
-                  model.modelLabel,
-                  evidenceConfigLabel(highLeader, strings, multiDbModel),
-                )}
-              >
-                {strings.evidenceHighDashboard}
-              </OverviewDashboardLink>
-            ) : null}
-          </div>
-        );
-      })}
+    <div data-testid="overview-baseline" data-hardware={props.member.hardware}>
+      <CellValue {...props} />
     </div>
   );
 }
 
 /**
- * Every remaining hardware, accounted for in one line: those with no exact @50
- * read and not ranked by the secondary block, each with its explicit reason. The
- * standing note "not ranked does not mean slower" lives once in the methodology.
+ * One candidate platform's cell: its own read, then the same-precision delta
+ * against B200, a mismatch note when both sides exist but cannot compare, and
+ * the @100 leader line only when the leader flips.
  */
-function NotRankedLine({
-  model,
-  strings,
-}: {
-  model: OverviewModelSummary;
-  strings: OverviewStrings;
-}) {
-  // Hardware already named by the secondary section — its ranked rows or its
-  // coverage 'measured' list — is not repeated here; dedup against the secondary's
-  // measured labels so each hardware surfaces in exactly one place.
-  const secondaryMeasured = new Set(model.secondary?.measuredHardware);
-  const entries = model.notRanked.filter((entry) => !secondaryMeasured.has(entry.hardwareLabel));
-  if (entries.length === 0) return null;
-  // ∞ marks missing/unavailable only — it never enters ranking or gap math, so it
-  // must never render with a percent. The reason rides in title/aria; the page
-  // footer carries the legend.
-  return (
-    <p className="flex flex-wrap gap-x-3 gap-y-0.5 text-xs text-muted-foreground">
-      {entries.map((entry) => {
-        const reason = notRankedReason(entry, strings);
-        return (
-          <span
-            key={entry.hardware}
-            title={reason}
-            aria-label={`${entry.hardwareLabel}: ${reason}`}
-          >
-            {entry.hardwareLabel} <span aria-hidden="true">∞</span>
-          </span>
-        );
-      })}
-    </p>
-  );
-}
-
-/** The full per-model hierarchy, identical on both surfaces. */
-function ModelHierarchy({
+function CandidateCell({
   locale,
   model,
+  pair,
   formatters,
   strings,
 }: {
   locale: OverviewLocale;
   model: OverviewModelSummary;
+  pair: OverviewHeadlinePairComparison;
   formatters: Formatters;
   strings: OverviewStrings;
 }) {
-  const cohorts = rankedCohorts(model.comparisonGroups);
+  const highLine = highLeadLine(pair, strings);
   return (
-    <div className="space-y-2 text-sm">
-      <PrimarySection
+    <div data-testid="overview-pair" data-pair={pair.id} className="space-y-1">
+      <CellValue
         locale={locale}
         model={model}
-        cohorts={cohorts}
+        member={pair.candidate}
         formatters={formatters}
         strings={strings}
       />
-      <SecondarySection locale={locale} model={model} formatters={formatters} strings={strings} />
-      <HighSection
-        locale={locale}
-        model={model}
-        cohorts={cohorts}
-        formatters={formatters}
-        strings={strings}
-      />
-      <NotRankedLine model={model} strings={strings} />
+      {pair.directDeltaPercent === null ? null : (
+        <p
+          data-testid="overview-pair-delta"
+          className="text-xs font-medium tabular-nums text-foreground"
+        >
+          {strings.delta(signedPercent(formatters, pair.directDeltaPercent))}
+        </p>
+      )}
+      {pair.deltaUnavailableReason === null ? null : (
+        <p data-testid="overview-pair-mismatch" className="text-xs text-muted-foreground">
+          {pair.deltaUnavailableReason === 'precision_mismatch'
+            ? strings.precisionMismatch(
+                pair.candidate.precision ?? '',
+                pair.baseline.precision ?? '',
+              )
+            : strings.versionMismatch}
+        </p>
+      )}
+      {highLine === null ? null : <p className="text-xs text-muted-foreground">{highLine}</p>}
     </div>
   );
 }
 
-/**
- * The model's name and — for a coverage-only model that shows no ranked chips —
- * its freshness. A ranked model dates each result on its own chip instead, so
- * the model-level date is only the fallback for a model with nothing to rank.
- */
-function ModelIdentity({
-  model,
-  formatters,
-}: {
-  model: OverviewModelSummary;
-  formatters: Formatters;
-}) {
-  const hasRankedResults =
-    rankedCohorts(model.comparisonGroups).length > 0 || model.secondary?.state === 'ranked';
-  const freshness =
-    hasRankedResults || model.latestWorkloadDate === null
-      ? null
-      : formatters.date(model.latestWorkloadDate);
-  return (
-    <div className="flex min-w-0 flex-wrap items-baseline gap-x-2 gap-y-1">
-      <h2 className="text-sm font-semibold leading-snug">{model.modelLabel}</h2>
-      {freshness ? (
-        <span className="text-xs text-muted-foreground tabular-nums">{freshness}</span>
-      ) : null}
-    </div>
-  );
+/** The model's display name, shared by both surfaces. */
+function ModelName({ model }: { model: OverviewModelSummary }) {
+  return <h2 className="text-sm font-semibold leading-snug">{model.modelLabel}</h2>;
 }
 
 interface SurfaceProps {
@@ -644,8 +329,10 @@ interface SurfaceProps {
   strings: OverviewStrings;
 }
 
-/** Model column plus one content column carrying the full hierarchy; no scroll. */
+/** Model column, the B200 reference, one column per candidate platform, details. */
 export function DesktopOverviewMatrix({ models, locale, formatters, strings }: SurfaceProps) {
+  const headlinePairs = models[0]?.headlinePairs ?? [];
+  const baselineHeader = headlinePairs[0]?.baseline.hardwareLabel ?? null;
   return (
     <div className="hidden xl:block">
       <table
@@ -654,8 +341,11 @@ export function DesktopOverviewMatrix({ models, locale, formatters, strings }: S
       >
         <caption className="sr-only">{strings.caption}</caption>
         <colgroup>
-          <col className="w-[18%]" />
-          <col className="w-[70%]" />
+          <col className="w-[17%]" />
+          {baselineHeader === null ? null : <col className="w-[13%]" />}
+          {headlinePairs.map((pair) => (
+            <col key={pair.id} className="w-[13.5%]" />
+          ))}
           <col className="w-[12%]" />
         </colgroup>
         <thead>
@@ -663,33 +353,54 @@ export function DesktopOverviewMatrix({ models, locale, formatters, strings }: S
             <th scope="col" className="px-4 py-2 text-left font-semibold lg:px-6">
               {strings.modelHeader}
             </th>
-            <th scope="col" className="px-4 py-2 text-left font-semibold">
-              {strings.resultsHeader}
-            </th>
+            {baselineHeader === null ? null : (
+              <th scope="col" className="px-4 py-2 text-left font-semibold">
+                {baselineHeader}
+              </th>
+            )}
+            {headlinePairs.map((pair) => (
+              <th key={pair.id} scope="col" className="px-4 py-2 text-left font-semibold">
+                {pair.candidate.hardwareLabel}
+              </th>
+            ))}
             <th scope="col" className="px-4 py-2 text-left font-semibold">
               {strings.detailsHeader}
             </th>
           </tr>
         </thead>
-        {models.map((model) => (
-          <tbody
-            key={model.model}
-            data-testid="overview-desktop-model"
-            data-model={model.model}
-            className="border-b border-border/50 last:border-b-0"
-          >
-            <tr className="align-top">
+        <tbody>
+          {models.map((model) => (
+            <tr
+              key={model.model}
+              data-testid="overview-desktop-model"
+              data-model={model.model}
+              className="border-b border-border/50 align-top last:border-b-0"
+            >
               <th scope="row" className="px-4 py-3 text-left align-top font-normal lg:px-6">
-                <ModelIdentity model={model} formatters={formatters} />
+                <ModelName model={model} />
               </th>
-              <td className="px-4 py-3">
-                <ModelHierarchy
-                  locale={locale}
-                  model={model}
-                  formatters={formatters}
-                  strings={strings}
-                />
-              </td>
+              {model.headlinePairs[0] === undefined ? null : (
+                <td className="px-4 py-3 align-top">
+                  <BaselineCell
+                    locale={locale}
+                    model={model}
+                    member={model.headlinePairs[0].baseline}
+                    formatters={formatters}
+                    strings={strings}
+                  />
+                </td>
+              )}
+              {model.headlinePairs.map((pair) => (
+                <td key={pair.id} className="px-4 py-3 align-top">
+                  <CandidateCell
+                    locale={locale}
+                    model={model}
+                    pair={pair}
+                    formatters={formatters}
+                    strings={strings}
+                  />
+                </td>
+              ))}
               <td className="px-4 py-3 align-top">
                 <OverviewDetailLink
                   href={detailHref(locale, model)}
@@ -700,14 +411,14 @@ export function DesktopOverviewMatrix({ models, locale, formatters, strings }: S
                 </OverviewDetailLink>
               </td>
             </tr>
-          </tbody>
-        ))}
+          ))}
+        </tbody>
       </table>
     </div>
   );
 }
 
-/** Below `xl`: the same hierarchy stacked as cards, always fully visible. */
+/** Below `xl`: the same cells stacked as one card per model, always fully visible. */
 export function MobileOverviewList({ models, locale, formatters, strings }: SurfaceProps) {
   return (
     <ul data-testid="overview-mobile-list" className="divide-y divide-border/50 xl:hidden">
@@ -716,15 +427,39 @@ export function MobileOverviewList({ models, locale, formatters, strings }: Surf
           <article
             data-testid="overview-mobile-model"
             data-model={model.model}
-            className="space-y-2 px-4 py-3"
+            className="space-y-2 px-4 py-3.5"
           >
-            <ModelIdentity model={model} formatters={formatters} />
-            <ModelHierarchy
-              locale={locale}
-              model={model}
-              formatters={formatters}
-              strings={strings}
-            />
+            <ModelName model={model} />
+            <div className="grid grid-cols-[84px_1fr] items-baseline gap-x-3 gap-y-1">
+              {model.headlinePairs[0] === undefined ? null : (
+                <>
+                  <span className="text-xs font-medium text-muted-foreground">
+                    {model.headlinePairs[0].baseline.hardwareLabel}
+                  </span>
+                  <BaselineCell
+                    locale={locale}
+                    model={model}
+                    member={model.headlinePairs[0].baseline}
+                    formatters={formatters}
+                    strings={strings}
+                  />
+                </>
+              )}
+              {model.headlinePairs.map((pair) => (
+                <div key={pair.id} className="contents">
+                  <span className="text-xs font-medium text-muted-foreground">
+                    {pair.candidate.hardwareLabel}
+                  </span>
+                  <CandidateCell
+                    locale={locale}
+                    model={model}
+                    pair={pair}
+                    formatters={formatters}
+                    strings={strings}
+                  />
+                </div>
+              ))}
+            </div>
             <OverviewDetailLink
               href={detailHref(locale, model)}
               model={model.model}
@@ -740,12 +475,59 @@ export function MobileOverviewList({ models, locale, formatters, strings }: Surf
   );
 }
 
+/**
+ * The service-level selector: one plain link per tier, so every view is a
+ * server-rendered, copyable URL — no client state. The displayed tier is
+ * inert text marked `aria-current`, never a self-link.
+ */
+export function OverviewTierSwitcher({
+  tier,
+  locale,
+  strings,
+}: {
+  tier: OverviewTier;
+  locale: OverviewLocale;
+  strings: OverviewStrings;
+}) {
+  const optionClass = 'px-3 py-1.5 tabular-nums';
+  return (
+    <nav
+      data-testid="overview-tier-switcher"
+      aria-label={strings.tierNavLabel}
+      className="mt-3 flex flex-wrap items-center gap-x-2 gap-y-1 text-xs"
+    >
+      <span className="text-muted-foreground">{strings.tierNavLabel}</span>
+      <div className="flex divide-x divide-border/60 overflow-hidden rounded-md border border-border/60">
+        {OVERVIEW_TIERS.map((option) =>
+          option === tier ? (
+            <span
+              key={option}
+              aria-current="page"
+              className={`${optionClass} bg-foreground font-semibold text-background`}
+            >
+              {option}
+            </span>
+          ) : (
+            <a
+              key={option}
+              href={overviewTierHref(locale, option)}
+              className={`${optionClass} text-muted-foreground transition-colors hover:bg-muted hover:text-foreground`}
+            >
+              {option}
+            </a>
+          ),
+        )}
+      </div>
+      <span className="text-muted-foreground">{strings.tierUnit}</span>
+    </nav>
+  );
+}
+
 export function OverviewMethodology({ strings }: { strings: OverviewStrings }) {
   return (
     <div className="space-y-1 border-t border-border/50 px-4 py-3 text-xs leading-snug text-muted-foreground lg:px-6">
       <p>{strings.methodologyNote}</p>
       <p>{strings.infinityLegend}</p>
-      <p>{strings.cohortNote}</p>
       <p>{strings.interpolationNote}</p>
     </div>
   );
diff --git a/packages/app/src/lib/overview-data.server.ts b/packages/app/src/lib/overview-data.server.ts
index a9d3cafa..0bfa5fe6 100644
--- a/packages/app/src/lib/overview-data.server.ts
+++ b/packages/app/src/lib/overview-data.server.ts
@@ -4,15 +4,25 @@ import { FIXTURES_MODE } from '@semianalysisai/inferencex-db/connection';
 import type { BenchmarkRow } from '@/lib/api';
 import { getCachedBenchmarks } from '@/lib/benchmark-data.server';
 import { DEFAULT_MODELS } from '@/lib/data-mappings';
-import { assembleOverviewPageData, type OverviewPageData } from '@/lib/overview-data';
+import {
+  assembleOverviewPageData,
+  OVERVIEW_PRIMARY_TIER,
+  type OverviewPageData,
+  type OverviewTier,
+} from '@/lib/overview-data';
 import { loadFixture } from '@/lib/test-fixtures';
 
-export async function getOverviewPageData(): Promise<OverviewPageData> {
+export async function getOverviewPageData(
+  tier: OverviewTier = OVERVIEW_PRIMARY_TIER,
+): Promise<OverviewPageData> {
   // E2E fixtures mode serves a small synthetic rows-by-display-model fixture
   // through the same assembler the live path uses, so a contract drift breaks
   // the fixture tests instead of silently stranding the page's data.
   if (FIXTURES_MODE) {
-    return assembleOverviewPageData(loadFixture<Record<string, BenchmarkRow[]>>('overview-rows'));
+    return assembleOverviewPageData(
+      loadFixture<Record<string, BenchmarkRow[]>>('overview-rows'),
+      tier,
+    );
   }
 
   // Fetch rows per db model, concatenated per display model (one display model
@@ -25,5 +35,5 @@ export async function getOverviewPageData(): Promise<OverviewPageData> {
     }),
   );
 
-  return assembleOverviewPageData(Object.fromEntries(entries));
+  return assembleOverviewPageData(Object.fromEntries(entries), tier);
 }
diff --git a/packages/app/src/lib/overview-data.test.ts b/packages/app/src/lib/overview-data.test.ts
index 96df49f9..29c49a45 100644
--- a/packages/app/src/lib/overview-data.test.ts
+++ b/packages/app/src/lib/overview-data.test.ts
@@ -3,15 +3,11 @@ import { describe, expect, it } from 'vitest';
 import overviewRowsFixture from '../../cypress/fixtures/api/overview-rows.json';
 
 import type { BenchmarkRow } from './api';
-import { getHardwareConfig } from './constants';
 import { DEFAULT_MODELS, Model, Precision } from './data-mappings';
-import { overviewConfigIdentityKey } from './overview-config-identity';
 import {
   assembleOverviewPageData,
-  buildOverviewHardwareOrder,
   buildOverviewModelSummary,
-  overviewPrimaryValue,
-  selectOverviewPrecision,
+  resolveOverviewTier,
   type OverviewModelSummary,
 } from './overview-data';
 
@@ -78,223 +74,63 @@ function frontierAt(
   );
 }
 
-/** One config's frontier at explicit [interactivity, throughput, date] knots. */
-function datedFrontier(
-  knots: [number, number, string][],
-  overrides: Partial<BenchmarkRow> = {},
-): BenchmarkRow[] {
-  return knots.map(([intvty, tput, date], index) =>
-    row({
-      conc: index + 1,
-      metrics: { median_intvty: intvty, output_tput_per_gpu: tput },
-      date,
-      ...overrides,
-    }),
-  );
+function headlinePairOf(summary: OverviewModelSummary, id: string) {
+  return summary.headlinePairs.find((pair) => pair.id === id);
 }
 
-function statusOf(summary: OverviewModelSummary, hardware: string) {
-  return summary.hardwareStatuses.find((status) => status.hardware === hardware);
-}
-
-/**
- * Leader of the model's first comparable cohort. Leadership only ever exists
- * inside a cohort, so every ranking assertion reads it from there — there is no
- * model-global winner to compare across engine families or deployment modes.
- */
-function primaryLeaderOf(summary: OverviewModelSummary) {
-  return summary.comparisonGroups[0]?.primaryRanking.leader ?? null;
-}
-
-describe('selectOverviewPrecision', () => {
-  it('returns null without any FP4/FP8 rows', () => {
-    expect(selectOverviewPrecision(Model.Qwen3_5, [row({ precision: Precision.BF16 })])).toBeNull();
-    expect(selectOverviewPrecision(Model.Qwen3_5, [row({ precision: Precision.INT4 })])).toBeNull();
-  });
-
-  it('ignores standard-decode rows when counting exact-@50 coverage', () => {
-    // FP8 has the most curves but every one is standard decode; FP4's single
-    // speculative curve is the only exact-@50 coverage, so it wins outnumbered.
-    const rows = [
-      ...frontier([1200, 1000, 800, 600], { hardware: 'b200', precision: Precision.FP4 }),
+describe('overview headline pairs', () => {
+  it('keeps each side on its own best precision and withholds the delta on a mismatch', () => {
+    const summary = buildOverviewModelSummary(Model.Qwen3_5, [
+      ...frontier([1200, 1000, 800, 600], {
+        hardware: 'mi355x',
+        precision: Precision.FP4,
+      }),
       ...frontier([1100, 900, 700, 500], {
-        hardware: 'gb200',
+        hardware: 'mi355x',
         precision: Precision.FP8,
-        spec_method: 'none',
       }),
       ...frontier([1000, 800, 600, 400], {
-        hardware: 'gb300',
+        hardware: 'b200',
         precision: Precision.FP8,
-        spec_method: 'none',
-      }),
-    ];
-
-    expect(selectOverviewPrecision(Model.Qwen3_5, rows)).toBe(Precision.FP4);
-  });
-});
-
-describe('buildOverviewModelSummary', () => {
-  it('keeps Kimi point releases as separate exact results, never one blended frontier', () => {
-    // kimik2.5 and kimik2.7-code both map to Model.Kimi_K2_5. Otherwise
-    // identical (same hardware/date/config), a model-blind key would merge them
-    // into one frontier — kimik2.5's 100 tok/s/user point would make the winner
-    // reachable there. Exact identity keeps kimik2.7-code's frontier pure.
-    const summary = buildOverviewModelSummary(Model.Kimi_K2_5, [
-      row({
-        model: 'kimik2.7-code',
-        conc: 1,
-        metrics: { median_intvty: 30, output_tput_per_gpu: 1200 },
-      }),
-      row({
-        model: 'kimik2.7-code',
-        conc: 2,
-        metrics: { median_intvty: 50, output_tput_per_gpu: 1000 },
       }),
-      row({
-        model: 'kimik2.7-code',
-        conc: 3,
-        metrics: { median_intvty: 75, output_tput_per_gpu: 800 },
-      }),
-      row({ model: 'kimik2.5', conc: 4, metrics: { median_intvty: 50, output_tput_per_gpu: 400 } }),
-      row({
-        model: 'kimik2.5',
-        conc: 5,
-        metrics: { median_intvty: 100, output_tput_per_gpu: 700 },
-      }),
-    ]);
-
-    const leader = primaryLeaderOf(summary);
-
-    expect(leader?.dbModel).toBe('kimik2.7-code');
-    expect(overviewPrimaryValue(leader!)).toBe(1000);
-    expect(leader?.tierValues.find(({ tier }) => tier === 100)).toEqual({
-      tier: 100,
-      value: null,
-      boundary: 'unreachable',
-      evidenceDate: null,
-    });
-  });
-
-  it('splits cohorts by db model so point releases never rank against each other', () => {
-    // Same hardware, engine group and deployment mode; only the db model
-    // differs, so the two point releases must land in separate cohorts.
-    const summary = buildOverviewModelSummary(Model.Kimi_K2_5, [
-      ...frontier([1200, 1000, 800, 600], { model: 'kimik2.5' }),
-      ...frontier([1100, 900, 700, 500], { model: 'kimik2.7-code' }),
-    ]);
-
-    expect(summary.comparisonGroups.map((group) => group.dbModel).toSorted()).toEqual([
-      'kimik2.5',
-      'kimik2.7-code',
-    ]);
-  });
-
-  it('never compares incompatible DeepSeek MTP engine families', () => {
-    const summary = buildOverviewModelSummary(Model.DeepSeek_V4_Pro, [
-      ...frontier([1200, 1000, 800, 600], { framework: 'dynamo-trt' }),
-      ...frontier([1100, 900, 700, 500], { framework: 'mori-sglang' }),
     ]);
 
-    expect(summary.comparisonGroups.map(({ engineGroup }) => engineGroup).toSorted()).toEqual([
-      'sglang',
-      'trt',
-    ]);
-    expect(summary.comparisonGroups.every((group) => group.primaryRanking.runnerUp === null)).toBe(
-      true,
-    );
+    const pair = headlinePairOf(summary, 'mi355x-vs-b200');
+    expect(pair?.candidate.read.value).toBe(1000);
+    expect(pair?.candidate.precision).toBe(Precision.FP4);
+    expect(pair?.baseline.read.value).toBe(800);
+    expect(pair?.baseline.precision).toBe(Precision.FP8);
+    expect(pair?.precision).toBeNull();
+    expect(pair?.directDeltaPercent).toBeNull();
+    expect(pair?.deltaUnavailableReason).toBe('precision_mismatch');
+    expect(pair?.highLeaderTransition).toBeNull();
   });
 
-  it('keeps aggregate and disaggregate denominators in separate cohorts', () => {
+  it('breaks equal exact-read coverage toward FP4', () => {
     const summary = buildOverviewModelSummary(Model.Qwen3_5, [
-      ...frontier([1200, 1000, 800, 600], { disagg: false }),
-      ...frontier([1800, 1500, 1200, 900], {
-        hardware: 'gb200',
-        disagg: true,
-        is_multinode: true,
+      ...frontier([1200, 1000, 800, 600], {
+        hardware: 'mi355x',
+        precision: Precision.FP4,
       }),
-    ]);
-    expect(summary.comparisonGroups.map(({ deploymentMode }) => deploymentMode).toSorted()).toEqual(
-      ['aggregated', 'disaggregated'],
-    );
-  });
-
-  it('counts aggregated GPUs as one pool and disaggregated as prefill + decode', () => {
-    const summary = buildOverviewModelSummary(Model.Qwen3_5, [
-      ...frontier([1200, 1000, 800, 600], { num_prefill_gpu: 4, num_decode_gpu: 4 }),
-      ...frontier([1800, 1500, 1200, 900], {
-        hardware: 'gb200',
-        disagg: true,
-        is_multinode: true,
-        num_prefill_gpu: 16,
-        num_decode_gpu: 16,
+      ...frontier([1100, 900, 700, 500], {
+        hardware: 'b200',
+        precision: Precision.FP4,
+      }),
+      ...frontier([1000, 800, 600, 400], {
+        hardware: 'mi355x',
+        precision: Precision.FP8,
+      }),
+      ...frontier([900, 700, 500, 300], {
+        hardware: 'b200',
+        precision: Precision.FP8,
       }),
     ]);
-    const leaderFor = (mode: string) =>
-      summary.comparisonGroups.find((group) => group.deploymentMode === mode)?.primaryRanking
-        .leader;
-
-    expect(leaderFor('aggregated')?.totalGpu).toBe(4);
-    expect(leaderFor('disaggregated')?.totalGpu).toBe(32);
-  });
-
-  it('ranks 50 and 100 independently across hardware with a real gap', () => {
-    // b200 leads 50 (1000 vs 800) but bottoms out at 100 (300), where gb200's
-    // shallower curve (400) takes over — two exact reads, so the gap is real.
-    const summary = buildOverviewModelSummary(Model.Qwen3_5, [
-      ...frontier([1200, 1000, 800, 300], { hardware: 'b200' }),
-      ...frontier([1100, 800, 600, 400], { hardware: 'gb200' }),
-    ]);
-    const [group] = summary.comparisonGroups;
-    const expectedHighConfigKey = overviewConfigIdentityKey(row({ hardware: 'gb200' }));
-
-    expect(group.primaryRanking.state).toBe('comparable');
-    expect(group.primaryRanking.leader?.hardware).toBe('b200');
-    expect(group.primaryRanking.runnerUp?.hardware).toBe('gb200');
-    expect(group.primaryRanking.gapPercent).toBeCloseTo((1000 / 800 - 1) * 100);
-    expect(group.highRanking.leader?.key).toBe(expectedHighConfigKey);
-    expect(group.highRanking.leader?.hardware).toBe('gb200');
-    expect(group.highLeaderTransition).toBe('changed_hardware');
-  });
-
-  it('picks the best per-hardware config independently at each tier', () => {
-    // On b200, sglang wins 50 (1000 > 700) while dynamo-trt wins 100 (400 > 200).
-    const summary = buildOverviewModelSummary(Model.Qwen3_5, [
-      ...frontier([1200, 1000, 800, 200], { framework: 'sglang' }),
-      ...frontier([1100, 700, 500, 400], { framework: 'dynamo-trt' }),
-    ]);
-    const [group] = summary.comparisonGroups;
-
-    expect(group.primaryRanking.leader).toMatchObject({ hardware: 'b200', framework: 'sglang' });
-    expect(group.highRanking.leader).toMatchObject({ hardware: 'b200', framework: 'dynamo-trt' });
-    expect(group.highLeaderTransition).toBe('same_hardware');
-  });
-
-  it('reports no_primary_baseline when 50 has no exact read but 100 does', () => {
-    // Frontier floor at 60 tok/s/user: tier 50 is clamped_low (no exact read),
-    // so the primary tier has no leader, while tier 100 sits inside [60, 110]
-    // and interpolates — a high leader with no primary baseline to compare to.
-    // FP4 rows so this insufficient-coverage precision is the primary (0-0 exact
-    // tie → FP4); the transition mechanics under test are precision-agnostic.
-    const summary = buildOverviewModelSummary(Model.Qwen3_5, [
-      ...frontierAt(
-        [
-          [60, 900],
-          [75, 800],
-          [90, 700],
-          [110, 600],
-        ],
-        { precision: Precision.FP4 },
-      ),
-    ]);
-    const [group] = summary.comparisonGroups;
 
-    expect(group.primaryRanking.state).toBe('insufficient_coverage');
-    expect(group.highLeaderTransition).toBe('no_primary_baseline');
+    expect(headlinePairOf(summary, 'mi355x-vs-b200')?.precision).toBe(Precision.FP4);
   });
 
-  it('never lets clamped or unreachable reads lead or form a gap at a tier', () => {
+  it('keeps an FP4 bucket and member boundaries when neither side has an exact read', () => {
     const summary = buildOverviewModelSummary(Model.Qwen3_5, [
-      // Frontier floor at 60 tok/s/user → 50 read is clamped_low.
       ...frontierAt(
         [
           [60, 900],
@@ -302,364 +138,310 @@ describe('buildOverviewModelSummary', () => {
           [80, 700],
           [90, 600],
         ],
-        { hardware: 'b300', precision: Precision.FP4 },
+        { hardware: 'mi355x', precision: Precision.FP4 },
       ),
-      // Frontier ceiling at 35 tok/s/user → 50 read is unreachable.
       ...frontierAt(
         [
           [20, 500],
-          [25, 450],
-          [30, 400],
-          [35, 350],
+          [30, 450],
+          [40, 400],
+          [45, 350],
         ],
         { hardware: 'b200', precision: Precision.FP4 },
       ),
     ]);
-    const [group] = summary.comparisonGroups;
 
-    expect(group.primaryRanking.state).toBe('insufficient_coverage');
-    expect(group.primaryRanking.gapPercent).toBeNull();
-    expect(group.primaryRanking.leader).toBeNull();
-    expect(group.hardwareStatuses.find((s) => s.hardware === 'b300')?.primary.boundary).toBe(
-      'clamped_low',
-    );
+    const pair = headlinePairOf(summary, 'mi355x-vs-b200');
+    expect(pair?.precision).toBeNull();
+    expect(pair?.candidate.precision).toBe(Precision.FP4);
+    expect(pair?.candidate.dbModel).toBe('qwen3.5');
+    expect(pair?.baseline.precision).toBe(Precision.FP4);
+    expect(pair?.candidate.read).toMatchObject({
+      value: null,
+      boundary: 'clamped_low',
+      config: { hardware: 'mi355x' },
+    });
+    expect(pair?.baseline.read).toMatchObject({
+      value: null,
+      boundary: 'unreachable',
+      config: { hardware: 'b200' },
+    });
+    expect(pair?.candidate.missingReason).toBe('no_exact_at_tier');
+    expect(pair?.baseline.missingReason).toBe('cannot_reach_at_tier');
   });
 
-  it('lets a lower exact read outrank a higher clamped one', () => {
-    // b300's clamped 900 is its frontier floor, not a 50 tok/s/user result, so
-    // gb200's real 500 leads and no gap is claimed against unmeasured coverage.
+  it('keeps an exact side and marks an unreachable side missing without a delta', () => {
     const summary = buildOverviewModelSummary(Model.Qwen3_5, [
+      ...frontier([1200, 1000, 800, 600], {
+        hardware: 'mi355x',
+        precision: Precision.FP4,
+      }),
       ...frontierAt(
         [
-          [60, 900],
-          [70, 800],
-          [80, 700],
-          [90, 600],
+          [20, 500],
+          [30, 450],
+          [40, 400],
+          [45, 350],
         ],
-        { hardware: 'b300' },
+        { hardware: 'b200', precision: Precision.FP4 },
       ),
-      ...frontier([700, 500, 400, 300], { hardware: 'gb200' }),
     ]);
-    const [group] = summary.comparisonGroups;
 
-    expect(group.primaryRanking.leader?.hardware).toBe('gb200');
-    expect(group.primaryRanking.state).toBe('single_measured');
-    expect(group.primaryRanking.gapPercent).toBeNull();
-    expect(group.hardwareStatuses.find((s) => s.hardware === 'b300')?.primary).toMatchObject({
-      value: 900,
-      boundary: 'clamped_low',
-    });
+    const pair = headlinePairOf(summary, 'mi355x-vs-b200');
+    expect(pair?.candidate.read).toMatchObject({ value: 1000, boundary: 'interpolated' });
+    expect(pair?.baseline.read).toMatchObject({ value: null, boundary: 'unreachable' });
+    expect(pair?.baseline.missingReason).toBe('cannot_reach_at_tier');
+    expect(pair?.directDeltaPercent).toBeNull();
   });
 
-  it('dates a model by its newest workload row, and null when it has none', () => {
+  it('selects member precisions independently and shares a pair precision only when they match', () => {
     const summary = buildOverviewModelSummary(Model.Qwen3_5, [
-      ...frontier([1200, 1000, 800, 600], { date: '2026-07-01' }),
-      ...frontier([1100, 900, 700, 500], { hardware: 'gb200', date: '2026-07-15' }),
+      ...frontier([1200, 1000, 800, 600], {
+        hardware: 'mi355x',
+        precision: Precision.FP8,
+      }),
+      ...frontier([1100, 900, 700, 500], {
+        hardware: 'b200',
+        precision: Precision.FP8,
+      }),
+      ...frontier([1300, 1100, 900, 700], {
+        hardware: 'gb300',
+        precision: Precision.FP4,
+      }),
+      ...frontier([1000, 800, 600, 400], {
+        hardware: 'b200',
+        precision: Precision.FP4,
+      }),
     ]);
 
-    expect(summary.latestWorkloadDate).toBe('2026-07-15');
-    expect(buildOverviewModelSummary(Model.Qwen3_5, []).latestWorkloadDate).toBeNull();
+    const matched = headlinePairOf(summary, 'mi355x-vs-b200');
+    expect(matched?.precision).toBe(Precision.FP8);
+    expect(matched?.directDeltaPercent).not.toBeNull();
+
+    const mismatched = headlinePairOf(summary, 'gb300-vs-b200');
+    expect(mismatched?.candidate.precision).toBe(Precision.FP4);
+    expect(mismatched?.baseline.precision).toBe(Precision.FP8);
+    expect(mismatched?.precision).toBeNull();
+    expect(mismatched?.deltaUnavailableReason).toBe('precision_mismatch');
   });
-});
 
-describe('overview precision selection and secondary', () => {
-  it('makes the wider exact-@50 coverage the primary precision', () => {
-    const summary = buildOverviewModelSummary(Model.Qwen3_5, [
-      ...frontier([1200, 1000, 800, 600], { hardware: 'b200', precision: Precision.FP4 }),
-      ...frontier([1100, 900, 700, 500], { hardware: 'gb200', precision: Precision.FP8 }),
-      ...frontier([1000, 800, 600, 400], { hardware: 'gb300', precision: Precision.FP8 }),
+  it('shows each release’s own read but never deltas across releases', () => {
+    const summary = buildOverviewModelSummary(Model.Kimi_K2_5, [
+      ...frontier([1200, 1000, 800, 600], {
+        model: 'kimik2.5',
+        hardware: 'b200',
+        precision: Precision.FP4,
+        date: '2026-07-10',
+      }),
+      ...frontier([1100, 900, 700, 500], {
+        model: 'kimik2.7-code',
+        hardware: 'mi355x',
+        precision: Precision.FP4,
+        date: '2026-07-20',
+      }),
     ]);
 
-    expect(summary.selectedPrecision).toBe(Precision.FP8);
+    const pair = headlinePairOf(summary, 'mi355x-vs-b200');
+    expect(pair?.candidate.read.config?.dbModel).toBe('kimik2.7-code');
+    expect(pair?.baseline.read.config?.dbModel).toBe('kimik2.5');
+    expect(pair?.dbModel).toBeNull();
+    expect(pair?.directDeltaPercent).toBeNull();
+    expect(pair?.deltaUnavailableReason).toBe('version_mismatch');
+    expect(pair?.highLeaderTransition).toBeNull();
   });
 
-  it('breaks an exact-@50 coverage tie toward FP4', () => {
+  it('computes signed candidate-relative-to-B200 deltas', () => {
     const summary = buildOverviewModelSummary(Model.Qwen3_5, [
-      ...frontier([1200, 1000, 800, 600], { hardware: 'b200', precision: Precision.FP4 }),
-      ...frontier([1100, 900, 700, 500], { hardware: 'gb200', precision: Precision.FP8 }),
+      ...frontier([1000, 800, 600, 400], {
+        hardware: 'mi355x',
+        precision: Precision.FP4,
+      }),
+      ...frontier([1200, 1000, 800, 600], {
+        hardware: 'b200',
+        precision: Precision.FP4,
+      }),
+      ...frontier([1400, 1200, 1000, 800], {
+        hardware: 'gb300',
+        precision: Precision.FP4,
+      }),
     ]);
 
-    expect(summary.selectedPrecision).toBe(Precision.FP4);
+    expect(headlinePairOf(summary, 'mi355x-vs-b200')?.directDeltaPercent).toBeCloseTo(-20);
+    expect(headlinePairOf(summary, 'gb300-vs-b200')?.directDeltaPercent).toBeCloseTo(20);
   });
 
-  it('ranks each precision within itself, never across FP4 and FP8', () => {
+  it('reports a hardware leader flip at @100 using independently best configs', () => {
     const summary = buildOverviewModelSummary(Model.Qwen3_5, [
-      ...frontier([1200, 1000, 800, 600], { hardware: 'b300', precision: Precision.FP4 }),
-      ...frontier([1100, 900, 700, 500], { hardware: 'b200', precision: Precision.FP4 }),
-      ...frontier([1000, 800, 600, 400], { hardware: 'gb200', precision: Precision.FP8 }),
+      ...frontier([1400, 1200, 800, 400], {
+        hardware: 'mi355x',
+        precision: Precision.FP4,
+        framework: 'sglang',
+      }),
+      ...frontier([1000, 700, 650, 600], {
+        hardware: 'mi355x',
+        precision: Precision.FP4,
+        framework: 'dynamo-trt',
+      }),
+      ...frontier([1200, 1000, 800, 650], {
+        hardware: 'b200',
+        precision: Precision.FP4,
+      }),
     ]);
 
-    // The FP4 cohort ranks only its own two hardware; the FP8 platform is the
-    // secondary precision, never a cross-precision delta in the primary cohort.
-    expect(summary.comparisonGroups[0].hardwareStatuses.map((s) => s.hardware).toSorted()).toEqual([
-      'b200',
-      'b300',
-    ]);
-    expect(summary.secondary?.precision).toBe(Precision.FP8);
+    const pair = headlinePairOf(summary, 'mi355x-vs-b200');
+    expect(pair?.candidate.read.config?.framework).toBe('sglang');
+    expect(pair?.candidate.highRead.config?.framework).toBe('dynamo-trt');
+    expect(pair?.highLeaderTransition).toBe('changed_hardware');
   });
 
-  it('keeps FP8 visible as secondary coverage when FP4 is primary', () => {
+  it('distinguishes standard-decode-only and unsupported-precision coverage per member', () => {
     const summary = buildOverviewModelSummary(Model.Qwen3_5, [
-      ...frontier([1200, 1000, 800, 600], { hardware: 'b300', precision: Precision.FP4 }),
-      ...frontier([1100, 900, 700, 500], { hardware: 'b200', precision: Precision.FP4 }),
-      ...frontier([1000, 800, 600, 400], { hardware: 'h200', precision: Precision.FP8 }),
+      ...frontier([1200, 1000, 800, 600], { hardware: 'b200', precision: Precision.FP4 }),
+      row({ hardware: 'mi355x', precision: Precision.FP8, spec_method: 'none' }),
+      row({ hardware: 'b300', precision: Precision.INT4 }),
     ]);
 
-    expect(summary.selectedPrecision).toBe(Precision.FP4);
-    expect(summary.secondary?.state).toBe('coverage');
-    expect(summary.secondary?.measuredHardware).toContain(getHardwareConfig('h200').label);
+    expect(headlinePairOf(summary, 'mi355x-vs-b200')?.candidate.missingReason).toBe(
+      'standard_decode_only',
+    );
+    expect(headlinePairOf(summary, 'b300-vs-b200')?.candidate.missingReason).toBe('int4_bf16_only');
   });
 
-  it('opens secondary rows only when FP8 ranks a pair on hardware FP4 lacks', () => {
-    const ranked = buildOverviewModelSummary(Model.Qwen3_5, [
-      ...frontier([1200, 1000, 800, 600], { hardware: 'b300', precision: Precision.FP4 }),
-      ...frontier([1150, 950, 750, 550], { hardware: 'b200', precision: Precision.FP4 }),
-      ...frontier([1100, 900, 700, 500], { hardware: 'gb300', precision: Precision.FP4 }),
-      ...frontier([1000, 800, 600, 400], { hardware: 'gb200', precision: Precision.FP8 }),
-      ...frontier([980, 780, 580, 380], { hardware: 'mi355x', precision: Precision.FP8 }),
-    ]);
-    // FP4 leads on three hardware; FP8 ranks a comparable pair on two it lacks.
-    expect(ranked.selectedPrecision).toBe(Precision.FP4);
-    expect(ranked.secondary?.state).toBe('ranked');
-
-    const covered = buildOverviewModelSummary(Model.Qwen3_5, [
-      ...frontier([1200, 1000, 800, 600], { hardware: 'gb200', precision: Precision.FP4 }),
-      ...frontier([1150, 950, 750, 550], { hardware: 'mi355x', precision: Precision.FP4 }),
-      ...frontier([900, 850, 650, 450], { hardware: 'gb200', precision: Precision.FP8 }),
-      ...frontier([880, 820, 620, 420], { hardware: 'mi355x', precision: Precision.FP8 }),
-    ]);
-    // Same FP8 pair, but both hardware already rank in FP4 → no new coverage.
-    expect(covered.secondary?.state).toBe('coverage');
-  });
-});
+  it('always returns all four fixed pairs for an empty model', () => {
+    const summary = buildOverviewModelSummary(Model.Qwen3_5, []);
 
-describe('overview not-ranked reasons', () => {
-  it('accounts for every page hardware with an exact value or one reason', () => {
-    const model = [
-      ...frontier([1200, 1000, 800, 600], { hardware: 'b300', precision: Precision.FP4 }),
-      ...frontier([1100, 900, 700, 500], { hardware: 'b200', precision: Precision.FP4 }),
-      // gb200 measured at the primary precision, but standard decode only.
-      ...frontier([1000, 800, 600, 400], {
-        hardware: 'gb200',
-        precision: Precision.FP4,
-        spec_method: 'none',
-      }),
-    ];
-    // mi355x reaches the page order only because another model measured it.
-    const other = frontier([900, 800, 700, 600], { hardware: 'mi355x', precision: Precision.FP4 });
-    const order = buildOverviewHardwareOrder([...model, ...other]);
-    const summary = buildOverviewModelSummary(Model.Qwen3_5, model, order);
-
-    // b300/b200 rank (exact @50) and are absent; gb200 and mi355x each get one.
-    expect(summary.notRanked.map((entry) => [entry.hardware, entry.reason])).toEqual([
-      ['gb200', 'standard_decode_only'],
-      ['mi355x', 'no_8k1k_data'],
+    expect(summary.headlinePairs.map(({ id }) => id)).toEqual([
+      'mi355x-vs-b200',
+      'b300-vs-b200',
+      'gb200-vs-b200',
+      'gb300-vs-b200',
     ]);
-  });
-
-  it('separates a frontier that tops out below 50 from a missing @50 read', () => {
-    const summary = buildOverviewModelSummary(Model.Qwen3_5, [
-      // b200 frontier ceiling at 45 tok/s/user → @50 unreachable.
-      ...frontierAt(
-        [
-          [20, 500],
-          [30, 450],
-          [40, 400],
-          [45, 350],
-        ],
-        { hardware: 'b200', precision: Precision.FP4 },
-      ),
-      // b300 frontier floor at 60 tok/s/user → @50 clamped_low, not unreachable.
-      ...frontierAt(
-        [
-          [60, 900],
-          [70, 800],
-          [80, 700],
-          [90, 600],
-        ],
-        { hardware: 'b300', precision: Precision.FP4 },
+    expect(summary.headlinePairs.every(({ precision }) => precision === null)).toBe(true);
+    expect(
+      summary.headlinePairs.every(
+        ({ candidate, baseline }) =>
+          candidate.missingReason === 'no_8k1k_data' && baseline.missingReason === 'no_8k1k_data',
       ),
-    ]);
-    const reasonOf = (hardware: string) =>
-      summary.notRanked.find((entry) => entry.hardware === hardware)?.reason;
-
-    expect(reasonOf('b200')).toBe('cannot_reach_at50');
-    expect(reasonOf('b300')).toBe('no_exact_at50');
+    ).toBe(true);
   });
 });
 
-describe('overview result-level evidence dates', () => {
-  it('dates each read from its own config frontier, never a sibling', () => {
-    const summary = buildOverviewModelSummary(Model.Qwen3_5, [
-      ...datedFrontier(
-        [
-          [40, 1100, '2026-07-10'],
-          [60, 900, '2026-07-10'],
-        ],
-        { hardware: 'b300', precision: Precision.FP4 },
-      ),
-      ...datedFrontier(
-        [
-          [40, 900, '2026-07-16'],
-          [60, 700, '2026-07-16'],
-        ],
-        { hardware: 'b200', precision: Precision.FP4 },
-      ),
-    ]);
-
-    expect(statusOf(summary, 'b300')?.primary.evidenceDate).toEqual({
-      from: '2026-07-10',
-      to: '2026-07-10',
-    });
-    expect(statusOf(summary, 'b200')?.primary.evidenceDate).toEqual({
-      from: '2026-07-16',
-      to: '2026-07-16',
-    });
+describe('tier-parameterized overview', () => {
+  it('resolves the tier query value and falls back to 50', () => {
+    expect(resolveOverviewTier('100')).toBe(100);
+    expect(resolveOverviewTier(['75', '30'])).toBe(75);
+    expect(resolveOverviewTier('40')).toBe(50);
+    expect(resolveOverviewTier('')).toBe(50);
+    expect(resolveOverviewTier(undefined)).toBe(50);
   });
 
-  it('collapses a same-day bracket and spans a cross-day one', () => {
-    const summary = buildOverviewModelSummary(Model.Qwen3_5, [
-      ...datedFrontier(
-        [
-          [40, 1100, '2026-07-10'],
-          [60, 900, '2026-07-10'],
-        ],
-        { hardware: 'b300', precision: Precision.FP4 },
-      ),
-      ...datedFrontier(
-        [
-          [40, 900, '2026-06-24'],
-          [60, 700, '2026-07-04'],
-        ],
-        { hardware: 'b200', precision: Precision.FP4 },
-      ),
-    ]);
-
-    expect(statusOf(summary, 'b300')?.primary.evidenceDate).toEqual({
-      from: '2026-07-10',
-      to: '2026-07-10',
-    });
-    expect(statusOf(summary, 'b200')?.primary.evidenceDate).toEqual({
-      from: '2026-06-24',
-      to: '2026-07-04',
-    });
+  it('stamps the displayed tier on the page and defaults to 50, down to empty models', () => {
+    expect(assembleOverviewPageData({}).tier).toBe(50);
+    const page = assembleOverviewPageData({}, 75);
+    expect(page.tier).toBe(75);
+    expect(page.models[0]?.headlinePairs[0]?.candidate.read.tier).toBe(75);
   });
 
-  it('ignores a hidden slower config even when it is newer', () => {
-    const summary = buildOverviewModelSummary(Model.Qwen3_5, [
-      ...datedFrontier(
-        [
-          [40, 1100, '2026-07-10'],
-          [60, 900, '2026-07-10'],
-        ],
-        { hardware: 'b300', precision: Precision.FP4, framework: 'sglang' },
-      ),
-      // Slower same-hardware config, newer runs — never backs the visible read.
-      ...datedFrontier(
-        [
-          [40, 700, '2026-07-20'],
-          [60, 500, '2026-07-20'],
+  it('reads every pair member and its delta at the requested tier', () => {
+    const page = assembleOverviewPageData(
+      {
+        [Model.Qwen3_5]: [
+          ...frontier([1000, 800, 600, 400], { hardware: 'mi355x', precision: Precision.FP4 }),
+          ...frontier([1200, 1000, 800, 600], { hardware: 'b200', precision: Precision.FP4 }),
         ],
-        { hardware: 'b300', precision: Precision.FP4, framework: 'dynamo-trt' },
-      ),
-    ]);
+      },
+      100,
+    );
 
-    expect(statusOf(summary, 'b300')?.primary.evidenceDate).toEqual({
-      from: '2026-07-10',
-      to: '2026-07-10',
-    });
+    const pair = headlinePairOf(
+      page.models.find((m) => m.model === Model.Qwen3_5)!,
+      'mi355x-vs-b200',
+    );
+    expect(pair?.candidate.read).toMatchObject({ tier: 100, value: 400 });
+    expect(pair?.baseline.read).toMatchObject({ tier: 100, value: 600 });
+    expect(pair?.directDeltaPercent).toBeCloseTo((400 / 600 - 1) * 100);
   });
 
-  it('dates @50 and @100 from their own bracketing knots', () => {
-    const summary = buildOverviewModelSummary(Model.Qwen3_5, [
-      ...datedFrontier(
+  it('turns an unreachable @50 side into an exact read on the 30 view', () => {
+    const rows = [
+      ...frontier([1200, 1000, 800, 600], { hardware: 'mi355x', precision: Precision.FP4 }),
+      ...frontierAt(
         [
-          [40, 1200, '2026-07-01'],
-          [60, 1000, '2026-07-01'],
-          [90, 700, '2026-07-15'],
-          [110, 500, '2026-07-15'],
+          [20, 500],
+          [30, 450],
+          [40, 400],
+          [45, 350],
         ],
-        { hardware: 'b300', precision: Precision.FP4 },
+        { hardware: 'b200', precision: Precision.FP4 },
       ),
-    ]);
-    const status = statusOf(summary, 'b300');
+    ];
+
+    const at50 = headlinePairOf(
+      assembleOverviewPageData({ [Model.Qwen3_5]: rows }).models.find(
+        (m) => m.model === Model.Qwen3_5,
+      )!,
+      'mi355x-vs-b200',
+    );
+    expect(at50?.baseline.missingReason).toBe('cannot_reach_at_tier');
 
-    expect(status?.primary.evidenceDate).toEqual({ from: '2026-07-01', to: '2026-07-01' });
-    expect(status?.high.evidenceDate).toEqual({ from: '2026-07-15', to: '2026-07-15' });
+    const at30 = headlinePairOf(
+      assembleOverviewPageData({ [Model.Qwen3_5]: rows }, 30).models.find(
+        (m) => m.model === Model.Qwen3_5,
+      )!,
+      'mi355x-vs-b200',
+    );
+    expect(at30?.baseline.read).toMatchObject({ tier: 30, value: 450 });
+    expect(at30?.baseline.missingReason).toBeNull();
+    expect(at30?.directDeltaPercent).toBeCloseTo((1200 / 450 - 1) * 100);
   });
-});
 
-describe('overview hardware coverage', () => {
-  it('holds hardware order fixed against input order and performance changes', () => {
-    const expected = ['gb300', 'gb200', 'b300', 'b200', 'mi355x'];
-    const build = (throughputs: number[], hardware: string[]) => {
-      const rows = hardware.flatMap((hw, index) =>
-        frontier(
-          [
-            throughputs[index] + 200,
-            throughputs[index],
-            throughputs[index] - 200,
-            throughputs[index] - 400,
+  it('re-selects each platform’s best bucket at the displayed tier', () => {
+    // FP4 wins @50 (1000 > 900) but tops out low; FP8 wins @100 (700 > 400).
+    const page = (tier?: 30 | 50 | 75 | 100) =>
+      assembleOverviewPageData(
+        {
+          [Model.Qwen3_5]: [
+            ...frontier([1200, 1000, 800, 400], { hardware: 'mi355x', precision: Precision.FP4 }),
+            ...frontier([1100, 900, 850, 700], { hardware: 'mi355x', precision: Precision.FP8 }),
+            ...frontier([1200, 1000, 800, 600], { hardware: 'b200', precision: Precision.FP8 }),
           ],
-          { hardware: hw },
-        ),
-      );
-      return buildOverviewModelSummary(Model.Qwen3_5, rows, buildOverviewHardwareOrder(rows));
-    };
-
-    // Same throughputs, hardware fed in the opposite order: the fastest platform
-    // swaps from mi355x to gb300 while the display order must not move.
-    const mi355xLeads = build([600, 700, 800, 900, 1000], expected);
-    const gb300Leads = build([600, 700, 800, 900, 1000], [...expected].toReversed());
-
-    expect(mi355xLeads.hardwareStatuses.map(({ hardware }) => hardware)).toEqual(expected);
-    expect(gb300Leads.hardwareStatuses.map(({ hardware }) => hardware)).toEqual(expected);
-    expect(primaryLeaderOf(mi355xLeads)).toMatchObject({ hardware: 'mi355x' });
-    expect(primaryLeaderOf(gb300Leads)).toMatchObject({ hardware: 'gb300' });
-    expect(gb300Leads.comparisonGroups[0].hardwareStatuses.map(({ hardware }) => hardware)).toEqual(
-      expected,
-    );
+        },
+        tier,
+      ).models.find((m) => m.model === Model.Qwen3_5)!;
+
+    const at50 = headlinePairOf(page(), 'mi355x-vs-b200');
+    expect(at50?.candidate.precision).toBe(Precision.FP4);
+    expect(at50?.deltaUnavailableReason).toBe('precision_mismatch');
+
+    const at100 = headlinePairOf(page(100), 'mi355x-vs-b200');
+    expect(at100?.candidate.precision).toBe(Precision.FP8);
+    expect(at100?.candidate.read.value).toBe(700);
+    expect(at100?.directDeltaPercent).toBeCloseTo((700 / 600 - 1) * 100);
   });
 
-  it('separates unsupported-precision coverage from hardware this model never ran', () => {
-    // Kimi shape: FP4 standard decode on B200, INT4 only on H200. H200 and
-    // MI355X reach the page order only because another model measured them.
-    const kimi = [
-      ...frontier([1200, 1000, 800, 600], {
-        model: 'kimik2.5',
-        hardware: 'b200',
+  it('never reports a leader transition on the 100 view', () => {
+    const rows = [
+      ...frontier([1400, 1200, 800, 400], {
+        hardware: 'mi355x',
         precision: Precision.FP4,
-        spec_method: 'none',
+        framework: 'sglang',
       }),
-      ...frontier([900, 800, 700, 600], {
-        model: 'kimik2.5',
-        hardware: 'h200',
-        precision: Precision.INT4,
-        spec_method: 'none',
+      ...frontier([1000, 700, 650, 600], {
+        hardware: 'mi355x',
+        precision: Precision.FP4,
+        framework: 'dynamo-trt',
       }),
+      ...frontier([1200, 1000, 800, 650], { hardware: 'b200', precision: Precision.FP4 }),
     ];
-    const otherModel = [
-      ...frontier([900, 800, 700, 600], { hardware: 'mi355x', precision: Precision.FP4 }),
-      ...frontier([700, 600, 500, 400], { hardware: 'h200', precision: Precision.FP8 }),
-    ];
-    const order = buildOverviewHardwareOrder([...kimi, ...otherModel]);
-    const summary = buildOverviewModelSummary(Model.Kimi_K2_5, kimi, order);
-
-    expect(order.map(({ hardware }) => hardware)).toEqual(['b200', 'mi355x', 'h200']);
-    expect(summary.hardwareStatuses.map(({ hardware }) => hardware)).toEqual([
-      'b200',
-      'mi355x',
-      'h200',
-    ]);
-    expect(statusOf(summary, 'b200')?.coverage.kind).toBe('standard_only');
-    expect(statusOf(summary, 'h200')?.coverage).toMatchObject({
-      kind: 'unsupported_precision_only',
-      availablePrecisions: [Precision.INT4],
-    });
-    expect(statusOf(summary, 'mi355x')?.coverage).toMatchObject({
-      kind: 'no_workload_data',
-      availablePrecisions: [],
-    });
+
+    const summaryAt = (tier?: 30 | 50 | 75 | 100) =>
+      assembleOverviewPageData({ [Model.Qwen3_5]: rows }, tier).models.find(
+        (m) => m.model === Model.Qwen3_5,
+      )!;
+    expect(headlinePairOf(summaryAt(), 'mi355x-vs-b200')?.highLeaderTransition).toBe(
+      'changed_hardware',
+    );
+    expect(headlinePairOf(summaryAt(100), 'mi355x-vs-b200')?.highLeaderTransition).toBeNull();
   });
 });
 
@@ -667,74 +449,67 @@ describe('overview hardware coverage', () => {
 // changes, the hand-built rows in overview-rows.json must be regenerated, and
 // this fails loudly instead of the e2e stranding empty data. Every expectation
 // is derived by running this same assembler over the fixture — never eyeballed —
-// so it locks the exact 8.7 states overview.cy.ts renders against.
+// so it locks the exact matrix cell states overview.cy.ts renders against.
 describe('assembleOverviewPageData over the overview-rows fixture', () => {
-  it('serves every 8.7 precision, coverage and evidence-date state through the real builder', () => {
+  it('serves every matrix cell state through the real builder', () => {
     const page = assembleOverviewPageData(
       overviewRowsFixture as unknown as Record<string, BenchmarkRow[]>,
     );
 
     expect(page.models).toHaveLength(DEFAULT_MODELS.size);
     expect(page.datasetThroughDate).toBe('2026-07-18');
-    expect(page.hardwareOrder.map((entry) => entry.hardware)).toEqual([
-      'gb300',
-      'gb200',
-      'b300',
-      'b200',
-      'mi355x',
-      'h200',
-      'mi325x',
-      'h100',
-    ]);
+    expect(page.tier).toBe(50);
+
+    // DeepSeek: FP4 exacts on B200/B300 carry the only same-precision delta with
+    // a cross-day evidence range; GB200's own best is FP8, so its value shows
+    // but the delta is withheld; MI355X and GB300 miss in opposite clamp
+    // directions.
+    const deepseek = page.models.find((m) => m.model === Model.DeepSeek_V4_Pro)!;
+    const dsB300 = headlinePairOf(deepseek, 'b300-vs-b200')!;
+    expect(dsB300.precision).toBe(Precision.FP4);
+    expect(dsB300.baseline.read.value).toBeCloseTo(900.219);
+    expect(dsB300.candidate.read.value).toBeCloseTo(1121.875);
+    expect(dsB300.candidate.read.evidenceDate).toEqual({ from: '2026-06-24', to: '2026-07-04' });
+    expect(dsB300.directDeltaPercent).toBeCloseTo(24.62, 1);
+    const dsGb200 = headlinePairOf(deepseek, 'gb200-vs-b200')!;
+    expect(dsGb200.candidate.precision).toBe(Precision.FP8);
+    expect(dsGb200.candidate.read.value).toBe(600);
+    expect(dsGb200.deltaUnavailableReason).toBe('precision_mismatch');
+    expect(headlinePairOf(deepseek, 'mi355x-vs-b200')?.candidate.missingReason).toBe(
+      'no_exact_at_tier',
+    );
+    expect(headlinePairOf(deepseek, 'gb300-vs-b200')?.candidate.missingReason).toBe(
+      'cannot_reach_at_tier',
+    );
 
-    // DeepSeek: FP4 primary ranks B300 over B200; FP8 stays visible as one
-    // coverage line (single measured hardware, not a second table); every other
-    // page hardware carries exactly one not-ranked reason, including BOTH clamp
-    // directions; and the leader's @50 read spans two calendar days — the only
-    // cross-day evidence range in the fixture, so the e2e can exercise the range
-    // label the current live dataset never produces.
-    const deepseek = page.models.find((m) => m.model === Model.DeepSeek_V4_Pro);
-    expect(deepseek?.selectedPrecision).toBe(Precision.FP4);
-    const dsCohort = deepseek?.comparisonGroups[0];
-    expect(dsCohort?.primaryRanking.state).toBe('comparable');
-    expect(dsCohort?.primaryRanking.leader?.hardware).toBe('b300');
-    expect(dsCohort?.highRanking.leader?.hardware).toBe('b200');
-    expect(
-      dsCohort?.hardwareStatuses.find((s) => s.hardware === 'b300')?.primary.evidenceDate,
-    ).toEqual({ from: '2026-06-24', to: '2026-07-04' });
-    expect(deepseek?.secondary?.state).toBe('coverage');
-    expect(deepseek?.secondary?.precision).toBe(Precision.FP8);
-    expect(deepseek?.secondary?.measuredHardware).toEqual(['GB200 NVL72']);
+    // MiniMax: only GB300 has workload rows, so its value stands alone — a
+    // missing baseline never yields a delta OR a mismatch note.
+    const minimax = page.models.find((m) => m.model === Model.MiniMax_M3)!;
+    const mmGb300 = headlinePairOf(minimax, 'gb300-vs-b200')!;
+    expect(mmGb300.baseline.missingReason).toBe('no_8k1k_data');
+    expect(mmGb300.candidate.read.value).toBe(700);
+    expect(mmGb300.directDeltaPercent).toBeNull();
+    expect(mmGb300.deltaUnavailableReason).toBeNull();
+
+    // Qwen: FP8 exacts on both sides yield the −16% delta and a leader flip at
+    // 100, while B300's FP4 best withholds its delta against the FP8 baseline.
+    const qwen = page.models.find((m) => m.model === Model.Qwen3_5)!;
+    const qwenMi = headlinePairOf(qwen, 'mi355x-vs-b200')!;
+    expect(qwenMi.precision).toBe(Precision.FP8);
+    expect(qwenMi.directDeltaPercent).toBeCloseTo(-15.56, 1);
+    expect(qwenMi.highLeaderTransition).toBe('changed_hardware');
+    const qwenB300 = headlinePairOf(qwen, 'b300-vs-b200')!;
+    expect(qwenB300.candidate.precision).toBe(Precision.FP4);
+    expect(qwenB300.candidate.read.value).toBeCloseTo(1150.625);
+    expect(qwenB300.deltaUnavailableReason).toBe('precision_mismatch');
+
+    // Kimi: no workload rows at all → every member of every pair says so.
+    const kimi = page.models.find((m) => m.model === Model.Kimi_K2_5)!;
     expect(
-      Object.fromEntries(
-        (deepseek?.notRanked ?? []).map((entry) => [entry.hardware, entry.reason]),
+      kimi.headlinePairs.every(
+        ({ candidate, baseline }) =>
+          candidate.missingReason === 'no_8k1k_data' && baseline.missingReason === 'no_8k1k_data',
       ),
-    ).toMatchObject({
-      gb300: 'cannot_reach_at50',
-      mi355x: 'no_exact_at50',
-      h100: 'standard_decode_only',
-      gb200: 'other_precision_only',
-      h200: 'no_8k1k_data',
-    });
-
-    // MiniMax: the wider exact-@50 FP8 coverage flips the primary precision to
-    // FP8 (the coverage rule), dropping FP4 to a single-hardware coverage line.
-    const minimax = page.models.find((m) => m.model === Model.MiniMax_M3);
-    expect(minimax?.selectedPrecision).toBe(Precision.FP8);
-    expect(minimax?.comparisonGroups[0]?.primaryRanking.leader?.hardware).toBe('h200');
-    expect(minimax?.secondary?.state).toBe('coverage');
-    expect(minimax?.secondary?.precision).toBe(Precision.FP4);
-    expect(minimax?.secondary?.measuredHardware).toEqual(['H200']);
-
-    // Qwen: FP4 primary, and FP8 secondary earns a FULL ranked block — it ranks a
-    // comparable pair AND adds MI355X, hardware FP4 has no exact @50 read for.
-    const qwen = page.models.find((m) => m.model === Model.Qwen3_5);
-    expect(qwen?.selectedPrecision).toBe(Precision.FP4);
-    expect(qwen?.secondary?.state).toBe('ranked');
-    expect(qwen?.secondary?.precision).toBe(Precision.FP8);
-    const qwenSecondaryCohort = qwen?.secondary?.comparisonGroups[0];
-    expect(qwenSecondaryCohort?.primaryRanking.state).toBe('comparable');
-    expect(qwenSecondaryCohort?.primaryRanking.leader?.hardware).toBe('b200');
-    expect(qwen?.secondary?.measuredHardware).toEqual(['B200', 'MI355X']);
+    ).toBe(true);
   });
 });
diff --git a/packages/app/src/lib/overview-data.ts b/packages/app/src/lib/overview-data.ts
index 6a4e4d59..17365c7d 100644
--- a/packages/app/src/lib/overview-data.ts
+++ b/packages/app/src/lib/overview-data.ts
@@ -5,24 +5,24 @@ import { parallelismLabel } from '@/components/inference/utils/parallelism-label
 import type { BenchmarkRow } from './api';
 import { buildAvailabilityHwKey } from './chart-utils';
 import { getHardwareConfig, getModelSortIndex } from './constants';
-import {
-  DEFAULT_MODELS,
-  getModelExclusion,
-  getModelLabel,
-  Precision,
-  type Model,
-} from './data-mappings';
-import { buildExclusion, type Exclusion } from './exclusion';
+import { DEFAULT_MODELS, getModelLabel, Precision, type Model } from './data-mappings';
 import { overviewConfigIdentityKey } from './overview-config-identity';
 import { computeTcoFeed, type TcoTierBoundary } from './tco-feed';
 
 export const OVERVIEW_WORKLOAD = { isl: 8192, osl: 1024 } as const;
 export const OVERVIEW_TIERS = [30, 50, 75, 100] as const;
-/** Headline service point every platform is ranked at. */
+export type OverviewTier = (typeof OVERVIEW_TIERS)[number];
+/** Default service point; ?tier= re-renders the matrix at any OVERVIEW_TIERS value. */
 export const OVERVIEW_PRIMARY_TIER = 50;
 /** High-interactivity capability read; also drives the high-tier leader transition. */
 export const OVERVIEW_HIGH_TIER = 100;
 
+/** The `?tier=` query value as a displayable tier; anything else falls back to 50. */
+export function resolveOverviewTier(raw: string | string[] | undefined): OverviewTier {
+  const candidate = Number(Array.isArray(raw) ? raw[0] : raw);
+  return OVERVIEW_TIERS.find((tier) => tier === candidate) ?? OVERVIEW_PRIMARY_TIER;
+}
+
 export interface OverviewTierValue {
   tier: number;
   value: number | null;
@@ -81,179 +81,84 @@ export interface OverviewTierRead {
 }
 
 /**
- * Why a hardware is or is not competing in the model's ranking. Coverage is
- * derived from the raw workload rows alone, so a hardware never disappears just
- * because a filter or a ranking excluded it:
+ * Why a pair member shows `∞` instead of a value, from most to least
+ * fundamental gap:
  *
- *  - `comparable_spec` — speculative results at the selected precision, ranked;
- *  - `standard_only` — selected precision measured, but standard decode only;
- *  - `alternate_precision_only` — measured at the other FP4/FP8 precision;
- *  - `other_engine_group_only` — speculative, but only in an engine
- *    comparability group no other hardware here shares;
- *  - `unsupported_precision_only` — measured only outside FP4/FP8;
- *  - `no_workload_data` — no 8K/1K single-turn row for this model.
+ *  - `no_8k1k_data` — no 8K/1K single-turn row for this hardware at all;
+ *  - `int4_bf16_only` — measured only outside FP4/FP8;
+ *  - `standard_decode_only` — FP4/FP8 measured, but standard decode only;
+ *  - `cannot_reach_at_tier` — speculative frontier tops out below the
+ *    displayed tier;
+ *  - `no_exact_at_tier` — no exact read at the displayed tier for any other
+ *    reason (an under-swept frontier that starts above the tier).
  */
-export type OverviewCoverageKind =
-  | 'comparable_spec'
-  | 'standard_only'
-  | 'alternate_precision_only'
-  | 'other_engine_group_only'
-  | 'unsupported_precision_only'
-  | 'no_workload_data';
-
-export interface OverviewHardwareCoverage {
-  hardware: string;
-  hardwareLabel: string;
-  kind: OverviewCoverageKind;
-  availablePrecisions: string[];
-}
+export type OverviewMissingReason =
+  | 'standard_decode_only'
+  | 'int4_bf16_only'
+  | 'no_8k1k_data'
+  | 'cannot_reach_at_tier'
+  | 'no_exact_at_tier';
 
-/** One hardware in the page-wide display order, which never encodes performance. */
-export interface OverviewHardwareOrderEntry {
-  hardware: string;
-  hardwareLabel: string;
-}
+export const OVERVIEW_HEADLINE_PAIR_DEFINITIONS = [
+  { id: 'mi355x-vs-b200', candidateHardware: 'mi355x', baselineHardware: 'b200' },
+  { id: 'b300-vs-b200', candidateHardware: 'b300', baselineHardware: 'b200' },
+  { id: 'gb200-vs-b200', candidateHardware: 'gb200', baselineHardware: 'b200' },
+  { id: 'gb300-vs-b200', candidateHardware: 'gb300', baselineHardware: 'b200' },
+] as const;
 
-export interface OverviewHardwareStatus {
+export type OverviewHeadlinePairId = (typeof OVERVIEW_HEADLINE_PAIR_DEFINITIONS)[number]['id'];
+
+export interface OverviewHeadlinePairMember {
   hardware: string;
   hardwareLabel: string;
-  coverage: OverviewHardwareCoverage;
   /**
-   * Primary-tier read of this hardware's best configuration in the surrounding
-   * config set — the same read the ranking used, so a row can never display a
-   * number the ranking did not. Null read when the hardware has no ranked
-   * configuration; `coverage.kind` says why.
+   * The precision of this platform's own best qualified speculative read at
+   * the displayed tier, selected independently per platform (exact first, then
+   * value, tie → FP4). Null when the platform has no FP4/FP8 speculative
+   * bucket at all.
    */
-  primary: OverviewTierRead;
-  /** High-tier read, taken from this hardware's own best config at that tier. */
-  high: OverviewTierRead;
-  isPrimaryLeader: boolean;
-  /** Signed percentage against the primary leader; null for the leader itself
-   * and whenever either side is not an exact in-range read. */
-  primaryDeltaPercent: number | null;
-}
-
-/** Whether a cohort's throughput denominator is aggregated or disaggregated. */
-export type OverviewDeploymentMode = 'aggregated' | 'disaggregated';
-
-/**
- * Ranking state at one tier: `comparable` (≥2 exact hardware reads, gap
- * computable), `single_measured` (one exact read), or `insufficient_coverage`
- * (no exact read — every read is clamped or unreachable).
- */
-export type OverviewRankingState = 'comparable' | 'single_measured' | 'insufficient_coverage';
-
-export interface OverviewTierRanking {
-  tier: number;
-  state: OverviewRankingState;
-  leader: OverviewConfigResult | null;
-  runnerUp: OverviewConfigResult | null;
-  gapPercent: number | null;
-}
-
-/**
- * One comparable cohort: exact configs that share an engine comparability group
- * (per the model's exclusion policy) and a deployment mode. Tiers are ranked
- * independently inside a cohort; configs never compare across cohorts.
- */
-export interface OverviewComparisonGroup {
-  id: string;
-  dbModel: string;
-  engineGroup: string | null;
-  deploymentMode: OverviewDeploymentMode;
-  hardwareStatuses: OverviewHardwareStatus[];
-  primaryRanking: OverviewTierRanking;
-  highRanking: OverviewTierRanking;
+  precision: string | null;
+  /** The dbModel (point release) backing this platform's selected read. */
+  dbModel: string | null;
+  /** Exact read when available; non-exact reads retain their boundary but never a numeric value. */
+  read: OverviewTierRead;
+  /** Best @100 read from this platform's own selected dbModel and precision. */
+  highRead: OverviewTierRead;
+  missingReason: OverviewMissingReason | null;
+}
+
+export interface OverviewHeadlinePairComparison {
+  id: OverviewHeadlinePairId;
+  label: string;
+  /** The precision shared by both displayed reads when the delta is comparable; null otherwise. */
+  precision: string | null;
+  /** The dbModel shared by both displayed reads when the delta is comparable; null otherwise. */
+  dbModel: string | null;
+  candidate: OverviewHeadlinePairMember;
+  baseline: OverviewHeadlinePairMember;
   /**
-   * How the high-tier leader relates to the primary-tier leader: a new hardware
-   * took the lead (`changed_hardware`), the same hardware held it
-   * (`same_hardware`), there was no primary leader to compare against
-   * (`no_primary_baseline`), or the high tier has no leader at all (`null`).
+   * Candidate relative to B200. Non-null only when both displayed-tier reads
+   * are exact, share one precision AND one dbModel, and B200 > 0 — a delta
+   * never mixes FP4 with FP8 or two point releases.
    */
-  highLeaderTransition: 'same_hardware' | 'changed_hardware' | 'no_primary_baseline' | null;
-}
-
-export interface OverviewSecondaryPrecision {
-  precision: string;
-  /** 'ranked' → full rows render; 'coverage' → one compact line renders. */
-  state: 'ranked' | 'coverage';
-  /** Populated when state === 'ranked'; [] otherwise. Built by the SAME pipeline as the primary (cohorts, rankings, transitions) over the secondary precision's rows. */
-  comparisonGroups: OverviewComparisonGroup[];
-  hardwareStatuses: OverviewHardwareStatus[];
-  /** Hardware labels measured at this precision (any decode mode), page order. Always populated. */
-  measuredHardware: string[];
-}
-
-/**
- * Why a page hardware carries no visible value for a model. Derived from the
- * existing coverage kinds and ranking outcomes, never a parallel truth:
- *
- *  - `standard_decode_only` — from `standard_only`;
- *  - `other_precision_only` — from `alternate_precision_only` and unranked by
- *    the secondary block; the renderer prints 'FP8 only'/'FP4 only';
- *  - `int4_bf16_only` — from `unsupported_precision_only`;
- *  - `different_serving_cohort` — from `other_engine_group_only`;
- *  - `no_8k1k_data` — from `no_workload_data`;
- *  - `cannot_reach_at50` — comparable spec whose frontier tops out below 50;
- *  - `no_exact_at50` — comparable spec with no exact @50 read for any other reason.
- */
-export type OverviewNotRankedReason =
-  | 'standard_decode_only'
-  | 'other_precision_only'
-  | 'int4_bf16_only'
-  | 'different_serving_cohort'
-  | 'no_8k1k_data'
-  | 'cannot_reach_at50'
-  | 'no_exact_at50';
-
-export interface OverviewNotRankedEntry {
-  hardware: string;
-  hardwareLabel: string;
-  reason: OverviewNotRankedReason;
-  /** FP4/FP8 precisions this hardware measured, for the 'FP8 only' style copy. */
-  precisions: string[];
+  directDeltaPercent: number | null;
+  /** Why two displayed exact reads still carry no delta; null when a delta exists or a side is missing. */
+  deltaUnavailableReason: 'precision_mismatch' | 'version_mismatch' | null;
+  highLeaderTransition: 'same_hardware' | 'changed_hardware' | null;
 }
 
 export interface OverviewModelSummary {
   model: Model;
   modelLabel: string;
-  /**
-   * Primary precision: whichever of FP4/FP8 covers more unique hardware with an
-   * exact @50 speculative read; a tie (including 0-0 when FP4/FP8 rows exist)
-   * goes to FP4. Null only when the model has no FP4/FP8 workload rows at all.
-   */
-  selectedPrecision: string | null;
-  /**
-   * Every page hardware's coverage and tier reads for this model, in the
-   * page-wide order. Leadership is claimed here only when a single cohort
-   * exists; with several cohorts the claim belongs to `comparisonGroups`.
-   */
-  hardwareStatuses: OverviewHardwareStatus[];
-  /** Comparable cohorts, each ranked independently at the primary and high tiers. */
-  comparisonGroups: OverviewComparisonGroup[];
-  /** The other of FP4/FP8, or null when it has no workload rows. Ranks only
-   * within itself — never against the primary precision. */
-  secondary: OverviewSecondaryPrecision | null;
-  /**
-   * One entry per page hardware with no visible value — neither an exact @50
-   * read in the primary block nor a ranked read in a `ranked` secondary. Every
-   * page hardware is therefore accounted for per model: an exact value, or one
-   * not-ranked reason.
-   */
-  notRanked: OverviewNotRankedEntry[];
-  /**
-   * Newest workload row measured for this model. Now that every visible result
-   * carries its own `evidenceDate`, this exists ONLY to date a coverage-only
-   * model with zero visible results. Null when the model has no workload row.
-   */
-  latestWorkloadDate: string | null;
-  emptyReason: 'no_8k1k_single_turn_data' | 'no_fp4_fp8_data' | null;
+  /** Fixed candidate-vs-B200 serving-stack comparisons, one per matrix column. */
+  headlinePairs: OverviewHeadlinePairComparison[];
 }
 
 export interface OverviewPageData {
   models: OverviewModelSummary[];
-  hardwareOrder: { hardware: string; hardwareLabel: string }[];
   datasetThroughDate: string | null;
+  /** The service point every headline pair was read at. */
+  tier: OverviewTier;
 }
 
 /** Precisions the overview may rank, in preference order. */
@@ -269,25 +174,6 @@ function overviewWorkloadRows(rows: readonly BenchmarkRow[]): BenchmarkRow[] {
   );
 }
 
-/**
- * Page-wide hardware display order: every base hardware with FP4/FP8 coverage
- * at the overview workload, in hardware-registry order. Built before precision
- * selection and before any ranking, so neither a precision fallback nor a
- * measurement change can reorder or drop a column.
- */
-export function buildOverviewHardwareOrder(
-  rows: readonly BenchmarkRow[],
-): OverviewHardwareOrderEntry[] {
-  const hardware = new Set(
-    overviewWorkloadRows(rows)
-      .filter((row) => OVERVIEW_PRECISIONS.includes(row.precision))
-      .map((row) => row.hardware),
-  );
-  return [...hardware]
-    .toSorted((a, b) => getModelSortIndex(a) - getModelSortIndex(b) || a.localeCompare(b))
-    .map((entry) => ({ hardware: entry, hardwareLabel: getHardwareConfig(entry).label }));
-}
-
 /**
  * Newest raw workload row across the page. Deliberately not derived from the
  * retained winners: a precision or engine the page never ranks still dates the
@@ -300,14 +186,6 @@ export function overviewDatasetThroughDate(rows: readonly BenchmarkRow[]): strin
   );
 }
 
-/** Newest date among a model's own workload rows; null when it has none. */
-function latestWorkloadDateOf(workloadRows: readonly BenchmarkRow[]): string | null {
-  return workloadRows.reduce<string | null>(
-    (latest, row) => (latest === null || row.date > latest ? row.date : latest),
-    null,
-  );
-}
-
 /**
  * Speculative configs for one precision, grouped by exact deployment identity.
  * Standard-decode rows never enter the pool, so a precision is ranked on its
@@ -335,56 +213,6 @@ function buildPrecisionConfigs(
   return configs;
 }
 
-/** Unique hardware whose best @50 read is exact (in-range) among `configs`. */
-function exactPrimaryHardware(configs: readonly OverviewConfigResult[]): Set<string> {
-  return new Set(
-    [...readsByHardwareAtTier(configs, OVERVIEW_PRIMARY_TIER).values()]
-      .filter(isExactTierRead)
-      .map((read) => read.config.hardware),
-  );
-}
-
-/** Whichever precision covers more exact-@50 hardware; a tie (incl. 0-0) → FP4. */
-function selectPrimaryFrom(
-  configsByPrecision: ReadonlyMap<string, readonly OverviewConfigResult[]>,
-): string {
-  const exactCount = (precision: string) =>
-    exactPrimaryHardware(configsByPrecision.get(precision) ?? []).size;
-  return exactCount(Precision.FP8) > exactCount(Precision.FP4) ? Precision.FP8 : Precision.FP4;
-}
-
-/**
- * Primary precision by exact-@50 hardware coverage: FP4 and FP8 are built and
- * counted independently, the wider coverage wins, and a tie (including 0-0 when
- * FP4/FP8 rows exist) goes to FP4. Null only when no FP4/FP8 workload row exists
- * — coverage volume of one vendor's curves can shift which precision is primary,
- * but never which precisions are ranked.
- */
-export function selectOverviewPrecision(
-  model: Model,
-  workloadRows: readonly BenchmarkRow[],
-): string | null {
-  if (!workloadRows.some((row) => OVERVIEW_PRECISIONS.includes(row.precision))) return null;
-  const configsByPrecision = new Map(
-    OVERVIEW_PRECISIONS.map(
-      (precision) => [precision, buildPrecisionConfigs(model, workloadRows, precision)] as const,
-    ),
-  );
-  return selectPrimaryFrom(configsByPrecision);
-}
-
-function tierValueAt(config: OverviewConfigResult, tier: number): number | null {
-  return config.tierValues.find((value) => value.tier === tier)?.value ?? null;
-}
-
-export function overviewPrimaryValue(config: OverviewConfigResult): number | null {
-  return tierValueAt(config, OVERVIEW_PRIMARY_TIER);
-}
-
-export function overviewHighValue(config: OverviewConfigResult): number | null {
-  return tierValueAt(config, OVERVIEW_HIGH_TIER);
-}
-
 function readConfigAtTier(config: OverviewConfigResult, tier: number): OverviewTierRead {
   const tierValue = config.tierValues.find((value) => value.tier === tier);
   return {
@@ -420,8 +248,7 @@ function compareTierReads(a: ConfigTierRead, b: ConfigTierRead): number {
 /**
  * One read per hardware at `tier`: its best exact in-range read, or — only when
  * the hardware has no exact read at all — its best out-of-range read, kept so
- * the hardware still surfaces as a coverage gap. Ranking and hardware rows both
- * consume this, so a row can never display a number the ranking did not use.
+ * the hardware still surfaces as a coverage gap rather than disappearing.
  */
 function readsByHardwareAtTier(
   configs: readonly OverviewConfigResult[],
@@ -445,199 +272,188 @@ function nullTierRead(tier: number): OverviewTierRead {
   return { tier, value: null, boundary: null, evidenceDate: null, config: null };
 }
 
-/**
- * Classify one hardware's coverage from its raw workload rows. `comparable`
- * lists the hardware sharing an engine comparability group with at least one
- * other hardware; a platform whose only speculative results sit in an engine
- * family nobody else here uses is reported as coverage rather than a competitor.
- */
-function coverageKindFor(
-  hardwareRows: readonly BenchmarkRow[],
-  selectedPrecision: string | null,
-  hardware: string,
-  comparable: ReadonlySet<string>,
-): OverviewCoverageKind {
-  if (hardwareRows.length === 0) return 'no_workload_data';
-  const candidates = hardwareRows.filter((row) => OVERVIEW_PRECISIONS.includes(row.precision));
-  if (candidates.length === 0) return 'unsupported_precision_only';
-  const selected = candidates.filter((row) => row.precision === selectedPrecision);
-  if (selected.length === 0) return 'alternate_precision_only';
-  if (!selected.some((row) => row.spec_method !== 'none')) return 'standard_only';
-  return comparable.size === 0 || comparable.has(hardware)
-    ? 'comparable_spec'
-    : 'other_engine_group_only';
-}
-
-/**
- * One status row per entry of `coverage`, which already carries the stable
- * hardware order. `configs` scopes the tier reads and the leader: the whole
- * model's ranked set at model level, one cohort's set inside a comparison group.
- */
-function buildHardwareStatuses(
-  coverage: readonly OverviewHardwareCoverage[],
-  configs: readonly OverviewConfigResult[],
-  withLeader: boolean,
-): OverviewHardwareStatus[] {
-  const primaryReads = readsByHardwareAtTier(configs, OVERVIEW_PRIMARY_TIER);
-  const highReads = readsByHardwareAtTier(configs, OVERVIEW_HIGH_TIER);
-  const leader = withLeader
-    ? [...primaryReads.values()].filter(isExactTierRead).toSorted(compareTierReads)[0]
-    : undefined;
-
-  return coverage.map((hardwareCoverage) => {
-    const { hardware, hardwareLabel } = hardwareCoverage;
-    const primary = primaryReads.get(hardware) ?? nullTierRead(OVERVIEW_PRIMARY_TIER);
-    const isPrimaryLeader = leader !== undefined && primary.config?.key === leader.config.key;
-    return {
-      hardware,
-      hardwareLabel,
-      coverage: hardwareCoverage,
-      primary,
-      high: highReads.get(hardware) ?? nullTierRead(OVERVIEW_HIGH_TIER),
-      isPrimaryLeader,
-      primaryDeltaPercent:
-        leader !== undefined && !isPrimaryLeader && isExactTierRead(primary) && leader.value > 0
-          ? (primary.value / leader.value - 1) * 100
-          : null,
-    };
-  });
-}
-
-/**
- * Rank one comparable cohort at a single tier. Every configuration is read at
- * `tier`, each hardware contributes only its best exact read, and the surviving
- * hardware are ordered by that read — so each tier is ranked on its own
- * evidence and a 50 winner is never carried into the 100 ranking.
- */
-function rankAtTier(configs: readonly OverviewConfigResult[], tier: number): OverviewTierRanking {
-  const hardwareLeaders = [...readsByHardwareAtTier(configs, tier).values()]
-    .filter(isExactTierRead)
-    .toSorted(compareTierReads);
-
-  const [leader, runnerUp] = hardwareLeaders;
-  return {
-    tier,
-    state:
-      hardwareLeaders.length >= 2
-        ? 'comparable'
-        : hardwareLeaders.length === 1
-          ? 'single_measured'
-          : 'insufficient_coverage',
-    leader: leader?.config ?? null,
-    runnerUp: runnerUp?.config ?? null,
-    gapPercent:
-      leader && runnerUp && runnerUp.value > 0 ? (leader.value / runnerUp.value - 1) * 100 : null,
-  };
+function nonComparableAsMissing(
+  read: OverviewTierRead | undefined,
+  tier: number,
+): OverviewTierRead {
+  if (read === undefined) return nullTierRead(tier);
+  return isExactTierRead(read) ? read : { ...read, value: null, evidenceDate: null };
 }
 
-interface OverviewCohort {
+interface OverviewHeadlinePairBucket {
+  precision: string;
   dbModel: string;
-  engineGroup: string | null;
-  deploymentMode: OverviewDeploymentMode;
   configs: OverviewConfigResult[];
+  tierReads: Map<string, ConfigTierRead>;
+  exactCount: number;
+  newestEvidence: string;
 }
 
-/**
- * How a cohort's high-tier leader relates to its primary-tier leader. Compared
- * on hardware, not config identity: a same-hardware engine or precision swap
- * between tiers is not a hardware change, and a high tier with no primary
- * baseline to compare against says so rather than implying a change.
- */
-function computeHighLeaderTransition(
-  primaryRanking: OverviewTierRanking,
-  highRanking: OverviewTierRanking,
-): OverviewComparisonGroup['highLeaderTransition'] {
-  if (highRanking.leader === null) return null;
-  if (primaryRanking.leader === null) return 'no_primary_baseline';
-  return highRanking.leader.hardware === primaryRanking.leader.hardware
-    ? 'same_hardware'
-    : 'changed_hardware';
-}
+function buildHeadlinePairBuckets(
+  configsByPrecision: ReadonlyMap<string, readonly OverviewConfigResult[]>,
+  hardware: ReadonlySet<string>,
+  tier: OverviewTier,
+): OverviewHeadlinePairBucket[] {
+  const buckets: OverviewHeadlinePairBucket[] = [];
+  for (const precision of OVERVIEW_PRECISIONS) {
+    const byDbModel = new Map<string, OverviewConfigResult[]>();
+    for (const config of configsByPrecision.get(precision) ?? []) {
+      if (!hardware.has(config.hardware)) continue;
+      const configs = byDbModel.get(config.dbModel);
+      if (configs) configs.push(config);
+      else byDbModel.set(config.dbModel, [config]);
+    }
 
-function buildComparisonGroup(
-  id: string,
-  cohort: OverviewCohort,
-  coverage: readonly OverviewHardwareCoverage[],
-): OverviewComparisonGroup {
-  const cohortHardware = new Set(cohort.configs.map(({ hardware }) => hardware));
-  const primaryRanking = rankAtTier(cohort.configs, OVERVIEW_PRIMARY_TIER);
-  const highRanking = rankAtTier(cohort.configs, OVERVIEW_HIGH_TIER);
-  return {
-    id,
-    dbModel: cohort.dbModel,
-    engineGroup: cohort.engineGroup,
-    deploymentMode: cohort.deploymentMode,
-    hardwareStatuses: buildHardwareStatuses(
-      coverage.filter(({ hardware }) => cohortHardware.has(hardware)),
-      cohort.configs,
-      true,
-    ),
-    primaryRanking,
-    highRanking,
-    highLeaderTransition: computeHighLeaderTransition(primaryRanking, highRanking),
-  };
+    for (const [dbModel, configs] of byDbModel) {
+      const tierReads = readsByHardwareAtTier(configs, tier);
+      const exactReads = [...tierReads.values()].filter(isExactTierRead);
+      buckets.push({
+        precision,
+        dbModel,
+        configs,
+        tierReads,
+        exactCount: exactReads.length,
+        newestEvidence: exactReads.reduce(
+          (latest, read) =>
+            read.evidenceDate !== null && read.evidenceDate.to > latest
+              ? read.evidenceDate.to
+              : latest,
+          configs.reduce(
+            (latest, config) => (config.latestDate > latest ? config.latestDate : latest),
+            '',
+          ),
+        ),
+      });
+    }
+  }
+  return buckets;
 }
 
-/**
- * Partition configurations into cohorts that may actually be compared: only
- * configs sharing a db model (one display model spans several point releases,
- * which must never rank against each other), an engine comparability group (the
- * model's exclusion policy — DeepSeek MTP acceptance forcing differs per engine
- * family) and a throughput denominator (aggregated vs disaggregated) are ranked
- * against each other.
- */
-function buildCohorts(
-  exclusion: Exclusion,
-  configs: readonly OverviewConfigResult[],
-): Map<string, OverviewCohort> {
-  const cohorts = new Map<string, OverviewCohort>();
-  for (const config of configs) {
-    const engineGroup = exclusion.groupOf(config.hwKey);
-    const deploymentMode: OverviewDeploymentMode = config.disagg ? 'disaggregated' : 'aggregated';
-    const id = `${config.dbModel}|${engineGroup ?? 'any'}|${deploymentMode}`;
-    const cohort = cohorts.get(id);
-    if (cohort) cohort.configs.push(config);
-    else
-      cohorts.set(id, { dbModel: config.dbModel, engineGroup, deploymentMode, configs: [config] });
+function missingReasonForHeadlineMember(
+  workloadRows: readonly BenchmarkRow[],
+  hardware: string,
+  read: OverviewTierRead,
+): OverviewMissingReason | null {
+  if (isExactTierRead(read)) return null;
+  const hardwareRows = workloadRows.filter((row) => row.hardware === hardware);
+  if (hardwareRows.length === 0) return 'no_8k1k_data';
+  const supportedRows = hardwareRows.filter((row) => OVERVIEW_PRECISIONS.includes(row.precision));
+  if (supportedRows.length === 0) return 'int4_bf16_only';
+  if (!supportedRows.some((row) => row.spec_method !== 'none')) return 'standard_decode_only';
+  return read.boundary === 'unreachable' ? 'cannot_reach_at_tier' : 'no_exact_at_tier';
+}
+
+function headlineLeaderTransition(
+  candidatePrimary: OverviewTierRead,
+  baselinePrimary: OverviewTierRead,
+  candidateHigh: OverviewTierRead,
+  baselineHigh: OverviewTierRead,
+): OverviewHeadlinePairComparison['highLeaderTransition'] {
+  if (
+    !isExactTierRead(candidatePrimary) ||
+    !isExactTierRead(baselinePrimary) ||
+    !isExactTierRead(candidateHigh) ||
+    !isExactTierRead(baselineHigh) ||
+    candidatePrimary.value === baselinePrimary.value ||
+    candidateHigh.value === baselineHigh.value
+  ) {
+    return null;
   }
-  return cohorts;
+  const primaryLeaderIsCandidate = candidatePrimary.value > baselinePrimary.value;
+  const highLeaderIsCandidate = candidateHigh.value > baselineHigh.value;
+  return primaryLeaderIsCandidate === highLeaderIsCandidate ? 'same_hardware' : 'changed_hardware';
 }
 
-/**
- * Hardware sharing an engine comparability group with at least one other
- * platform. Partitioned by engine group alone, never by the ranking cohorts:
- * cohorts also split on deployment mode, so a platform measured only
- * disaggregated would look isolated while running the very engine family its
- * peers run — and `other_engine_group_only` would then assert a reason that is
- * false.
- */
-function comparableHardware(
-  exclusion: Exclusion,
-  configs: readonly OverviewConfigResult[],
-): Set<string> {
-  const byEngineGroup = new Map<string | null, Set<string>>();
-  for (const config of configs) {
-    const group = exclusion.groupOf(config.hwKey);
-    const groupHardware = byEngineGroup.get(group);
-    if (groupHardware) groupHardware.add(config.hardware);
-    else byEngineGroup.set(group, new Set([config.hardware]));
-  }
-  return new Set(
-    [...byEngineGroup.values()]
-      .filter((groupHardware) => groupHardware.size > 1)
-      .flatMap((groupHardware) => [...groupHardware]),
+function buildHeadlinePairs(
+  model: Model,
+  workloadRows: readonly BenchmarkRow[],
+  tier: OverviewTier,
+): OverviewHeadlinePairComparison[] {
+  // FP4 and FP8 configs are built independently and never blend within a read.
+  const configsByPrecision = new Map(
+    OVERVIEW_PRECISIONS.map(
+      (precision) => [precision, buildPrecisionConfigs(model, workloadRows, precision)] as const,
+    ),
   );
-}
+  // Each platform selects its own best qualified bucket (dbModel × precision):
+  // exact at the displayed tier first, then value, tie → FP4, newest evidence,
+  // lexical dbModel. One bucket per member keeps point releases from blending
+  // within a read.
+  const buildMember = (memberHardware: string): OverviewHeadlinePairMember => {
+    const best = buildHeadlinePairBuckets(configsByPrecision, new Set([memberHardware]), tier)
+      .map((memberBucket) => ({
+        bucket: memberBucket,
+        read: nonComparableAsMissing(memberBucket.tierReads.get(memberHardware), tier),
+      }))
+      .toSorted(
+        (a, b) =>
+          Number(isExactTierRead(b.read)) - Number(isExactTierRead(a.read)) ||
+          (b.read.value ?? -1) - (a.read.value ?? -1) ||
+          OVERVIEW_PRECISIONS.indexOf(a.bucket.precision) -
+            OVERVIEW_PRECISIONS.indexOf(b.bucket.precision) ||
+          b.bucket.newestEvidence.localeCompare(a.bucket.newestEvidence) ||
+          a.bucket.dbModel.localeCompare(b.bucket.dbModel),
+      )[0];
+    const primary = best?.read ?? nullTierRead(tier);
+    const high = nonComparableAsMissing(
+      best === undefined
+        ? undefined
+        : readsByHardwareAtTier(best.bucket.configs, OVERVIEW_HIGH_TIER).get(memberHardware),
+      OVERVIEW_HIGH_TIER,
+    );
+    return {
+      hardware: memberHardware,
+      hardwareLabel: getHardwareConfig(memberHardware, model).label,
+      precision: best?.bucket.precision ?? null,
+      dbModel: best?.bucket.dbModel ?? null,
+      read: primary,
+      highRead: high,
+      missingReason: missingReasonForHeadlineMember(workloadRows, memberHardware, primary),
+    };
+  };
 
-function buildComparisonGroups(
-  cohorts: ReadonlyMap<string, OverviewCohort>,
-  coverage: readonly OverviewHardwareCoverage[],
-): OverviewComparisonGroup[] {
-  const leadValue = (group: OverviewComparisonGroup): number =>
-    group.primaryRanking.leader ? (overviewPrimaryValue(group.primaryRanking.leader) ?? -1) : -1;
-  return [...cohorts.entries()]
-    .map(([id, cohort]) => buildComparisonGroup(id, cohort, coverage))
-    .toSorted((a, b) => leadValue(b) - leadValue(a) || a.id.localeCompare(b.id));
+  return OVERVIEW_HEADLINE_PAIR_DEFINITIONS.map((definition) => {
+    const candidate = buildMember(definition.candidateHardware);
+    const baseline = buildMember(definition.baselineHardware);
+    const bothExact = isExactTierRead(candidate.read) && isExactTierRead(baseline.read);
+    const comparable =
+      bothExact &&
+      candidate.precision === baseline.precision &&
+      candidate.dbModel === baseline.dbModel;
+    return {
+      id: definition.id,
+      label: `${candidate.hardwareLabel} vs ${baseline.hardwareLabel}`,
+      precision: comparable ? candidate.precision : null,
+      dbModel: comparable ? candidate.dbModel : null,
+      candidate,
+      baseline,
+      directDeltaPercent:
+        comparable &&
+        candidate.read.value !== null &&
+        baseline.read.value !== null &&
+        baseline.read.value > 0
+          ? (candidate.read.value / baseline.read.value - 1) * 100
+          : null,
+      deltaUnavailableReason: bothExact
+        ? candidate.precision === baseline.precision
+          ? candidate.dbModel === baseline.dbModel
+            ? null
+            : 'version_mismatch'
+          : 'precision_mismatch'
+        : null,
+      // On the 100 view the displayed reads ARE the high reads, so a
+      // transition line would restate the cell values.
+      highLeaderTransition:
+        comparable && tier !== OVERVIEW_HIGH_TIER
+          ? headlineLeaderTransition(
+              candidate.read,
+              baseline.read,
+              candidate.highRead,
+              baseline.highRead,
+            )
+          : null,
+    };
+  });
 }
 
 function buildConfigResult(
@@ -706,264 +522,33 @@ function buildConfigResult(
   };
 }
 
-/**
- * Coverage for every hardware in `order`, derived from raw workload rows only.
- * Hardware the model never measured stays in the list as `no_workload_data`,
- * so a platform is never silently absent from one model's row.
- */
-function buildCoverage(
-  order: readonly OverviewHardwareOrderEntry[],
-  workloadRows: readonly BenchmarkRow[],
-  selectedPrecision: string | null,
-  comparable: ReadonlySet<string>,
-): OverviewHardwareCoverage[] {
-  const rowsByHardware = new Map<string, BenchmarkRow[]>();
-  for (const row of workloadRows) {
-    const hardwareRows = rowsByHardware.get(row.hardware);
-    if (hardwareRows) hardwareRows.push(row);
-    else rowsByHardware.set(row.hardware, [row]);
-  }
-
-  return order.map(({ hardware, hardwareLabel }) => {
-    const hardwareRows = rowsByHardware.get(hardware) ?? [];
-    return {
-      hardware,
-      hardwareLabel,
-      kind: coverageKindFor(hardwareRows, selectedPrecision, hardware, comparable),
-      availablePrecisions: [...new Set(hardwareRows.map((row) => row.precision))].toSorted(),
-    };
-  });
-}
-
-interface OverviewPrecisionBlock {
-  comparisonGroups: OverviewComparisonGroup[];
-  hardwareStatuses: OverviewHardwareStatus[];
-}
-
-/**
- * Rank one precision's speculative configs: cohorts, per-tier rankings and the
- * page-wide hardware statuses. The primary and secondary precisions are both
- * built through here, so neither can rank against the other.
- */
-function buildPrecisionBlock(
-  model: Model,
-  precision: string,
-  configs: readonly OverviewConfigResult[],
-  hardwareOrder: readonly OverviewHardwareOrderEntry[],
-  workloadRows: readonly BenchmarkRow[],
-): OverviewPrecisionBlock {
-  const exclusion = buildExclusion(getModelExclusion(model));
-  const cohorts = buildCohorts(exclusion, configs);
-  const comparable = comparableHardware(exclusion, configs);
-  const coverage = buildCoverage(hardwareOrder, workloadRows, precision, comparable);
-  const comparisonGroups = buildComparisonGroups(cohorts, coverage);
-  // Leadership is a claim against a comparable cohort. Across several cohorts
-  // there is no model-global leader, so the model row asserts none.
-  return {
-    comparisonGroups,
-    hardwareStatuses: buildHardwareStatuses(coverage, configs, comparisonGroups.length <= 1),
-  };
-}
-
-/**
- * The other of FP4/FP8. `ranked` (full rows) iff it both ranks a comparable
- * cohort AND covers hardware the primary precision lacks; otherwise `coverage`
- * (one compact line, ranked rows suppressed). `measuredHardware` always lists
- * every hardware measured at this precision so it never disappears from the page.
- */
-function buildSecondaryPrecision(
-  model: Model,
-  precision: string,
-  configs: readonly OverviewConfigResult[],
-  primaryExact: ReadonlySet<string>,
-  hardwareOrder: readonly OverviewHardwareOrderEntry[],
-  workloadRows: readonly BenchmarkRow[],
-): OverviewSecondaryPrecision | null {
-  const measuredSet = new Set(
-    workloadRows.filter((row) => row.precision === precision).map((row) => row.hardware),
-  );
-  if (measuredSet.size === 0) return null;
-  const measuredHardware = hardwareOrder
-    .filter(({ hardware }) => measuredSet.has(hardware))
-    .map(({ hardwareLabel }) => hardwareLabel);
-
-  const block = buildPrecisionBlock(model, precision, configs, hardwareOrder, workloadRows);
-  const addsCoverage = [...exactPrimaryHardware(configs)].some((hw) => !primaryExact.has(hw));
-  const hasComparable = block.comparisonGroups.some(
-    (group) => group.primaryRanking.state === 'comparable',
-  );
-  const ranked = hasComparable && addsCoverage;
-  return {
-    precision,
-    state: ranked ? 'ranked' : 'coverage',
-    comparisonGroups: ranked ? block.comparisonGroups : [],
-    hardwareStatuses: ranked ? block.hardwareStatuses : [],
-    measuredHardware,
-  };
-}
-
-/** Map one non-ranked hardware's coverage and read to its not-ranked reason. */
-function notRankedReasonFor(status: OverviewHardwareStatus): OverviewNotRankedReason {
-  switch (status.coverage.kind) {
-    case 'standard_only': {
-      return 'standard_decode_only';
-    }
-    case 'alternate_precision_only': {
-      return 'other_precision_only';
-    }
-    case 'unsupported_precision_only': {
-      return 'int4_bf16_only';
-    }
-    case 'other_engine_group_only': {
-      return 'different_serving_cohort';
-    }
-    case 'no_workload_data': {
-      return 'no_8k1k_data';
-    }
-    case 'comparable_spec': {
-      // Comparable-spec hardware reaches here only without an exact @50 read.
-      // Classify by INTERACTIVITY direction, not the boundary name: 'unreachable'
-      // = tier 50 above the frontier's max, i.e. even the fastest measured point
-      // is < 50 tok/s/user (the config is too slow, ever) → cannot_reach_at50.
-      // 'clamped_low' = tier below the frontier's min, i.e. every measured point
-      // is > 50 tok/s/user — an under-swept sweep gap more concurrency would
-      // cross, NOT incapability → no_exact_at50.
-      return status.primary.boundary === 'unreachable' ? 'cannot_reach_at50' : 'no_exact_at50';
-    }
-  }
-}
-
-/**
- * One not-ranked entry per page hardware with no visible value — no exact @50
- * read in the primary block and no ranked read in a `ranked` secondary. Reads
- * the same coverage/status machinery the blocks use, so it is a derived view
- * rather than a parallel source of truth.
- */
-function buildNotRanked(
-  primaryStatuses: readonly OverviewHardwareStatus[],
-  secondary: OverviewSecondaryPrecision | null,
-): OverviewNotRankedEntry[] {
-  const secondaryRanked =
-    secondary?.state === 'ranked'
-      ? new Set(
-          secondary.hardwareStatuses
-            .filter((status) => isExactTierRead(status.primary))
-            .map((status) => status.hardware),
-        )
-      : new Set<string>();
-
-  const entries: OverviewNotRankedEntry[] = [];
-  for (const status of primaryStatuses) {
-    if (isExactTierRead(status.primary) || secondaryRanked.has(status.hardware)) continue;
-    entries.push({
-      hardware: status.hardware,
-      hardwareLabel: status.hardwareLabel,
-      reason: notRankedReasonFor(status),
-      precisions: status.coverage.availablePrecisions.filter((p) =>
-        OVERVIEW_PRECISIONS.includes(p),
-      ),
-    });
-  }
-  return entries;
-}
-
-function emptyModelSummary(
-  model: Model,
-  emptyReason: NonNullable<OverviewModelSummary['emptyReason']>,
-  order: readonly OverviewHardwareOrderEntry[],
-  workloadRows: readonly BenchmarkRow[],
-): OverviewModelSummary {
-  const hardwareStatuses = buildHardwareStatuses(
-    buildCoverage(order, workloadRows, null, new Set()),
-    [],
-    false,
-  );
-  return {
-    model,
-    modelLabel: getModelLabel(model),
-    selectedPrecision: null,
-    hardwareStatuses,
-    comparisonGroups: [],
-    secondary: null,
-    notRanked: buildNotRanked(hardwareStatuses, null),
-    latestWorkloadDate: latestWorkloadDateOf(workloadRows),
-    emptyReason,
-  };
-}
-
 export function buildOverviewModelSummary(
   model: Model,
   rows: BenchmarkRow[],
-  hardwareOrder: readonly OverviewHardwareOrderEntry[] = buildOverviewHardwareOrder(rows),
+  tier: OverviewTier = OVERVIEW_PRIMARY_TIER,
 ): OverviewModelSummary {
-  const workloadRows = overviewWorkloadRows(rows);
-  if (workloadRows.length === 0) {
-    return emptyModelSummary(model, 'no_8k1k_single_turn_data', hardwareOrder, workloadRows);
-  }
-
-  if (!workloadRows.some((row) => OVERVIEW_PRECISIONS.includes(row.precision))) {
-    return emptyModelSummary(model, 'no_fp4_fp8_data', hardwareOrder, workloadRows);
-  }
-
-  // FP4 and FP8 speculative configs, built once and reused for selection, the
-  // primary block and the secondary block — the two precisions never blend.
-  const configsByPrecision = new Map(
-    OVERVIEW_PRECISIONS.map(
-      (precision) => [precision, buildPrecisionConfigs(model, workloadRows, precision)] as const,
-    ),
-  );
-  const selectedPrecision = selectPrimaryFrom(configsByPrecision);
-  const secondaryPrecision = OVERVIEW_PRECISIONS.find((p) => p !== selectedPrecision)!;
-  const primaryConfigs = configsByPrecision.get(selectedPrecision)!;
-
-  const primary = buildPrecisionBlock(
-    model,
-    selectedPrecision,
-    primaryConfigs,
-    hardwareOrder,
-    workloadRows,
-  );
-  const secondary = buildSecondaryPrecision(
-    model,
-    secondaryPrecision,
-    configsByPrecision.get(secondaryPrecision)!,
-    exactPrimaryHardware(primaryConfigs),
-    hardwareOrder,
-    workloadRows,
-  );
-
   return {
     model,
     modelLabel: getModelLabel(model),
-    selectedPrecision,
-    hardwareStatuses: primary.hardwareStatuses,
-    comparisonGroups: primary.comparisonGroups,
-    secondary,
-    notRanked: buildNotRanked(primary.hardwareStatuses, secondary),
-    latestWorkloadDate: latestWorkloadDateOf(workloadRows),
-    emptyReason: null,
+    headlinePairs: buildHeadlinePairs(model, overviewWorkloadRows(rows), tier),
   };
 }
 
 /**
  * Assemble the whole page from rows already grouped by DISPLAY model. Iterating
  * DEFAULT_MODELS fixes the output order and renders every active model — a model
- * with no rows falls through to its coverage-only states, exactly as the live
- * path does. The page-wide hardware order and dataset date come from the union
- * of every model's rows. The live server path and the e2e fixture path feed this
- * same function; only the row source differs.
+ * with no rows still gets its four pairs, every member carrying a missing
+ * reason. The live server path and the e2e fixture path feed this same
+ * function; only the row source differs.
  */
 export function assembleOverviewPageData(
   rowsByModel: Record<string, BenchmarkRow[]>,
+  tier: OverviewTier = OVERVIEW_PRIMARY_TIER,
 ): OverviewPageData {
   const perModel = [...DEFAULT_MODELS].map((model) => ({ model, rows: rowsByModel[model] ?? [] }));
-  const allRows = perModel.flatMap(({ rows }) => rows);
-  const hardwareOrder = buildOverviewHardwareOrder(allRows);
   return {
-    models: perModel.map(({ model, rows }) =>
-      buildOverviewModelSummary(model, rows, hardwareOrder),
-    ),
-    hardwareOrder,
-    datasetThroughDate: overviewDatasetThroughDate(allRows),
+    models: perModel.map(({ model, rows }) => buildOverviewModelSummary(model, rows, tier)),
+    datasetThroughDate: overviewDatasetThroughDate(perModel.flatMap(({ rows }) => rows)),
+    tier,
   };
 }
diff --git a/packages/app/src/lib/overview-links.test.ts b/packages/app/src/lib/overview-links.test.ts
index 0f6c4f35..c509a140 100644
--- a/packages/app/src/lib/overview-links.test.ts
+++ b/packages/app/src/lib/overview-links.test.ts
@@ -2,7 +2,7 @@ import { describe, expect, it } from 'vitest';
 
 import { Model, Precision } from './data-mappings';
 import type { OverviewConfigResult, OverviewModelSummary } from './overview-data';
-import { buildOverviewDashboardHref } from './overview-links';
+import { buildOverviewDashboardHref, detailHref } from './overview-links';
 
 const RUN_URL = 'https://github.com/SemiAnalysisAI/InferenceX/actions/runs/26714221123';
 
@@ -43,13 +43,7 @@ function summary(overrides: Partial<OverviewModelSummary> = {}): OverviewModelSu
   return {
     model: Model.Qwen3_5,
     modelLabel: 'Qwen 3.5',
-    selectedPrecision: Precision.FP4,
-    hardwareStatuses: [],
-    comparisonGroups: [],
-    secondary: null,
-    notRanked: [],
-    latestWorkloadDate: null,
-    emptyReason: null,
+    headlinePairs: [],
     ...overrides,
   };
 }
@@ -98,3 +92,11 @@ describe('buildOverviewDashboardHref', () => {
     );
   });
 });
+
+describe('detailHref', () => {
+  it('keeps the model drilldown precision-neutral because headline pairs may differ', () => {
+    expect(detailHref('en', summary())).toBe(
+      '/inference?g_model=Qwen-3.5-397B-A17B&i_seq=8k%2F1k&i_optimal=1',
+    );
+  });
+});
diff --git a/packages/app/src/lib/overview-links.ts b/packages/app/src/lib/overview-links.ts
index 8e9deb5e..132def11 100644
--- a/packages/app/src/lib/overview-links.ts
+++ b/packages/app/src/lib/overview-links.ts
@@ -6,7 +6,12 @@
  */
 
 import { runIdFromRunUrl } from './known-issues';
-import type { OverviewConfigResult, OverviewModelSummary } from './overview-data';
+import {
+  OVERVIEW_PRIMARY_TIER,
+  type OverviewConfigResult,
+  type OverviewModelSummary,
+  type OverviewTier,
+} from './overview-data';
 import type { UrlStateParams } from './url-state';
 
 /** The single-turn 8K-in/1K-out workload every overview link filters to. */
@@ -75,15 +80,24 @@ export function buildOverviewDashboardHref(
 }
 
 /**
- * Model-level dashboard view: the workload and precision this page ranked, with
- * no configuration pinned. Cohort-level evidence links narrow it further.
+ * Model-level dashboard view: precision-neutral, because the two headline pairs
+ * may select different precisions. Result-level evidence links narrow further.
  */
 export function detailHref(locale: 'en' | 'zh', model: OverviewModelSummary): string {
   const query = new URLSearchParams({
     g_model: model.model,
     i_seq: OVERVIEW_WORKLOAD_SEQ,
-    ...(model.selectedPrecision ? { i_prec: model.selectedPrecision } : {}),
     i_optimal: '1',
   });
   return `${inferenceRoute(locale)}?${query}`;
 }
+
+/**
+ * The overview itself at another service point. The default tier keeps the bare
+ * canonical URL; other tiers ride a plain `?tier=` query the server re-renders
+ * from, so any view is a copyable link.
+ */
+export function overviewTierHref(locale: 'en' | 'zh', tier: OverviewTier): string {
+  const base = locale === 'zh' ? '/zh/overview' : '/overview';
+  return tier === OVERVIEW_PRIMARY_TIER ? base : `${base}?tier=${tier}`;
+}
diff --git a/packages/app/src/lib/tco-feed.test.ts b/packages/app/src/lib/tco-feed.test.ts
index 4775c4d9..d6cf6203 100644
--- a/packages/app/src/lib/tco-feed.test.ts
+++ b/packages/app/src/lib/tco-feed.test.ts
@@ -180,6 +180,34 @@ describe('computeTcoFeed — frontier reads', () => {
     expect(rows[0].oldest_frontier_date).toBe('2026-04-25');
   });
 
+  it("uses only an exact interior frontier knot's date as evidence", () => {
+    const rows = computeTcoFeed(
+      [
+        makeRow({ itl: 1 / 20, otput: 1000, date: '2026-04-25' }),
+        makeRow({ itl: 1 / 50, otput: 400, date: '2026-05-20' }),
+        makeRow({ itl: 1 / 100, otput: 100, date: '2026-07-12' }),
+      ],
+      WORKLOAD_8K1K,
+      [50],
+    );
+
+    expect(rows[0].evidence_date).toEqual({ from: '2026-05-20', to: '2026-05-20' });
+  });
+
+  it('uses both bracketing frontier-knot dates as evidence between knots', () => {
+    const rows = computeTcoFeed(
+      [
+        makeRow({ itl: 1 / 20, otput: 1000, date: '2026-04-25' }),
+        makeRow({ itl: 1 / 50, otput: 400, date: '2026-05-20' }),
+        makeRow({ itl: 1 / 100, otput: 100, date: '2026-07-12' }),
+      ],
+      WORKLOAD_8K1K,
+      [75],
+    );
+
+    expect(rows[0].evidence_date).toEqual({ from: '2026-05-20', to: '2026-07-12' });
+  });
+
   it('groups by workload and hardware, sorted by hardware key', () => {
     const rows = computeTcoFeed(
       [
diff --git a/packages/app/src/lib/tco-feed.ts b/packages/app/src/lib/tco-feed.ts
index 9041b130..083c37d6 100644
--- a/packages/app/src/lib/tco-feed.ts
+++ b/packages/app/src/lib/tco-feed.ts
@@ -221,6 +221,7 @@ function bracketKnots(frontier: FrontierPoint[], tier: number): [FrontierPoint,
   if (tier >= frontier[last].interactivity) return [frontier[last], frontier[last]];
   let lo = 0;
   for (let i = 0; i < last; i++) {
+    if (frontier[i].interactivity === tier) return [frontier[i], frontier[i]];
     if (frontier[i].interactivity <= tier) lo = i;
   }
   return [frontier[lo], frontier[lo + 1]];

From 517e2cecf5877cc647329d8bb04db646f2585db1 Mon Sep 17 00:00:00 2001
From: Wenyao Gao <wgao11@u.rochester.edu>
Date: Wed, 22 Jul 2026 19:01:49 -0700
Subject: [PATCH 04/13] refactor(overview): drop dead dashboard-link helper and
 unreachable date guard

---
 .../overview/overview-detail-link.tsx         | 49 ++-----------------
 .../overview/overview-scorecard.tsx           |  7 ++-
 2 files changed, 7 insertions(+), 49 deletions(-)

diff --git a/packages/app/src/components/overview/overview-detail-link.tsx b/packages/app/src/components/overview/overview-detail-link.tsx
index f7b23298..2e1e44cb 100644
--- a/packages/app/src/components/overview/overview-detail-link.tsx
+++ b/packages/app/src/components/overview/overview-detail-link.tsx
@@ -5,14 +5,12 @@ import type { ReactNode } from 'react';
 
 import { track } from '@/lib/analytics';
 import { cn } from '@/lib/utils';
-import type { OverviewConfigResult, OverviewModelSummary } from '@/lib/overview-data';
-import { buildOverviewDashboardHref } from '@/lib/overview-links';
 
 /**
- * Model-level dashboard link. A plain anchor for the same reason the evidence
- * links below are: the dashboard reads its filters from a snapshot `url-state.ts`
- * takes at module evaluation, so a client-side `<Link>` navigation would land on
- * an unfiltered dashboard while the label promises the model's own view.
+ * Model-level dashboard link. A plain anchor, not a `<Link>`: the dashboard
+ * reads its filters from a snapshot `url-state.ts` takes at module evaluation,
+ * so a client-side navigation would land on an unfiltered dashboard while the
+ * label promises the model's own view.
  */
 export function OverviewDetailLink({
   href,
@@ -45,42 +43,3 @@ export function OverviewDetailLink({
     </a>
   );
 }
-
-const EVIDENCE_LINK_CLASS =
-  'group inline-flex min-h-11 w-fit items-center gap-1 whitespace-nowrap rounded-sm text-xs text-muted-foreground underline decoration-dotted underline-offset-4 transition-colors hover:text-foreground focus-visible:outline-none focus-visible:ring-[3px] focus-visible:ring-ring/50 motion-reduce:transition-none';
-
-/**
- * One evidence link: the /inference dashboard pre-filtered to a ranked
- * configuration. The visible text stays short so the compact rows scan; the
- * full deployment topology rides in `ariaLabel`, which the caller also makes
- * distinct per cohort — so no two evidence links share the accessible name
- * "Open filtered dashboard".
- *
- * A plain anchor, not a `<Link>`: the dashboard reads its share-link params
- * from a snapshot `url-state.ts` takes at module evaluation, so a client-side
- * navigation would arrive with every filter dropped — silently unfiltered while
- * the accessible name promises filtered evidence.
- */
-export function OverviewDashboardLink({
-  locale,
-  model,
-  config,
-  ariaLabel,
-  children,
-}: {
-  locale: 'en' | 'zh';
-  model: OverviewModelSummary;
-  config: OverviewConfigResult;
-  ariaLabel: string;
-  children: ReactNode;
-}) {
-  return (
-    <a
-      href={buildOverviewDashboardHref(locale, model, config)}
-      aria-label={ariaLabel}
-      className={EVIDENCE_LINK_CLASS}
-    >
-      {children}
-    </a>
-  );
-}
diff --git a/packages/app/src/components/overview/overview-scorecard.tsx b/packages/app/src/components/overview/overview-scorecard.tsx
index b5d09806..a6adb504 100644
--- a/packages/app/src/components/overview/overview-scorecard.tsx
+++ b/packages/app/src/components/overview/overview-scorecard.tsx
@@ -121,13 +121,12 @@ export function overviewFormatters(locale: OverviewLocale): Formatters {
 /**
  * A result's evidence dates, localized: a single day (`Jul 6` / `7月6日`) when
  * both backing frontier points share a date, else an en-dash range
- * (`Jun 24–Jul 4`). Null when the value has no backing date.
+ * (`Jun 24–Jul 4`). The caller renders nothing when there is no backing date.
  */
 function formatEvidenceDate(
   formatters: Formatters,
-  evidenceDate: { from: string; to: string } | null,
-): string | null {
-  if (evidenceDate === null) return null;
+  evidenceDate: { from: string; to: string },
+): string {
   const from = formatters.shortDate(evidenceDate.from);
   return evidenceDate.from === evidenceDate.to
     ? from

From c87d8e8ed6974abdba229f529929bfddf763a635 Mon Sep 17 00:00:00 2001
From: Wenyao Gao <wgao11@u.rochester.edu>
Date: Wed, 22 Jul 2026 23:11:38 -0700
Subject: [PATCH 05/13] refactor(overview): trim unused config fields; 44px
 minecraft audio targets

---
 .../minecraft/minecraft-toggles.tsx           |  2 +-
 packages/app/src/lib/overview-data.ts         | 37 -------------------
 packages/app/src/lib/overview-links.test.ts   |  9 -----
 3 files changed, 1 insertion(+), 47 deletions(-)

diff --git a/packages/app/src/components/minecraft/minecraft-toggles.tsx b/packages/app/src/components/minecraft/minecraft-toggles.tsx
index ccea82b6..073a0bff 100644
--- a/packages/app/src/components/minecraft/minecraft-toggles.tsx
+++ b/packages/app/src/components/minecraft/minecraft-toggles.tsx
@@ -6,7 +6,7 @@ import { track } from '@/lib/analytics';
 import { cn } from '@/lib/utils';
 
 const toggleClasses = cn(
-  'inline-flex items-center justify-center rounded-md p-2',
+  'inline-flex items-center justify-center rounded-md size-11',
   'text-muted-foreground hover:text-foreground hover:bg-accent',
   'transition-colors duration-200',
   'focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-ring focus-visible:ring-offset-2',
diff --git a/packages/app/src/lib/overview-data.ts b/packages/app/src/lib/overview-data.ts
index 17365c7d..1e1f9563 100644
--- a/packages/app/src/lib/overview-data.ts
+++ b/packages/app/src/lib/overview-data.ts
@@ -1,7 +1,5 @@
 import { resolveFrameworkPartLabel } from '@semianalysisai/inferencex-constants';
 
-import { parallelismLabel } from '@/components/inference/utils/parallelism-label';
-
 import type { BenchmarkRow } from './api';
 import { buildAvailabilityHwKey } from './chart-utils';
 import { getHardwareConfig, getModelSortIndex } from './constants';
@@ -47,7 +45,6 @@ export interface OverviewConfigResult {
   key: string;
   dbModel: string;
   hardware: string;
-  hardwareLabel: string;
   hwKey: string;
   framework: string;
   frameworkLabel: string;
@@ -55,19 +52,9 @@ export interface OverviewConfigResult {
   specLabel: string;
   disagg: boolean;
   precision: string;
-  offloadMode: string;
-  isMultinode: boolean;
-  numPrefillGpu: number;
-  numDecodeGpu: number;
-  /** Physical GPUs backing this config: one shared pool when aggregated, prefill
-   * plus decode when disaggregated. */
-  totalGpu: number;
-  parallelism: string;
-  image: string | null;
   sourceRunUrls: string[];
   tierValues: OverviewTierValue[];
   latestDate: string;
-  oldestFrontierDate: string;
 }
 
 /** One hardware's frontier read at a single tier, with its backing config. */
@@ -467,7 +454,6 @@ function buildConfigResult(
 
   const first = rows[0];
   const { hardware, framework, spec_method: specMethod, disagg } = first;
-  const latestRow = rows.reduce((a, b) => (b.date > a.date ? b : a));
   const sourceRunUrls = [
     ...new Set(rows.flatMap((row) => (row.run_url === null ? [] : [row.run_url]))),
   ].toSorted();
@@ -475,7 +461,6 @@ function buildConfigResult(
     key,
     dbModel: first.model,
     hardware,
-    hardwareLabel: getHardwareConfig(hardware, model).label,
     hwKey: buildAvailabilityHwKey(hardware, framework, specMethod, disagg),
     framework,
     frameworkLabel: resolveFrameworkPartLabel(model, framework),
@@ -483,27 +468,6 @@ function buildConfigResult(
     specLabel: resolveFrameworkPartLabel(model, specMethod),
     disagg,
     precision,
-    offloadMode: first.offload_mode ?? 'off',
-    isMultinode: first.is_multinode,
-    numPrefillGpu: first.num_prefill_gpu,
-    numDecodeGpu: first.num_decode_gpu,
-    totalGpu: disagg ? first.num_prefill_gpu + first.num_decode_gpu : first.num_decode_gpu,
-    parallelism: parallelismLabel({
-      tp: first.decode_tp,
-      ep: first.decode_ep,
-      dpAttention: first.decode_dp_attention,
-      disagg: first.disagg,
-      isMultinode: first.is_multinode,
-      prefillTp: first.prefill_tp,
-      prefillEp: first.prefill_ep,
-      prefillDpAttention: first.prefill_dp_attention,
-      prefillNumWorkers: first.prefill_num_workers,
-      decodeTp: first.decode_tp,
-      decodeEp: first.decode_ep,
-      decodeDpAttention: first.decode_dp_attention,
-      decodeNumWorkers: first.decode_num_workers,
-    }),
-    image: latestRow.image,
     sourceRunUrls,
     tierValues: feed.map((row) => {
       const value =
@@ -518,7 +482,6 @@ function buildConfigResult(
       };
     }),
     latestDate: feed[0].latest_date,
-    oldestFrontierDate: feed[0].oldest_frontier_date,
   };
 }
 
diff --git a/packages/app/src/lib/overview-links.test.ts b/packages/app/src/lib/overview-links.test.ts
index c509a140..71406dd6 100644
--- a/packages/app/src/lib/overview-links.test.ts
+++ b/packages/app/src/lib/overview-links.test.ts
@@ -16,7 +16,6 @@ function config(overrides: Partial<OverviewConfigResult> = {}): OverviewConfigRe
     key: 'qwen3.5|b200|sglang|mtp|agg|fp4',
     dbModel: 'qwen3.5',
     hardware: 'b200',
-    hardwareLabel: 'B200',
     hwKey: 'b200_sglang_mtp',
     framework: 'sglang',
     frameworkLabel: 'SGLang',
@@ -24,17 +23,9 @@ function config(overrides: Partial<OverviewConfigResult> = {}): OverviewConfigRe
     specLabel: 'MTP',
     disagg: false,
     precision: Precision.FP4,
-    offloadMode: 'off',
-    isMultinode: false,
-    numPrefillGpu: 8,
-    numDecodeGpu: 8,
-    totalGpu: 8,
-    parallelism: 'TP8',
-    image: 'inferencex/sglang:v0.5.0',
     sourceRunUrls: [RUN_URL],
     tierValues: [{ tier: 50, value: 1000, boundary: 'interpolated', evidenceDate: null }],
     latestDate: '2026-07-18',
-    oldestFrontierDate: '2026-07-11',
     ...overrides,
   };
 }

From 73d843e1cae8ed1e2756ba0785660eb856f3f239 Mon Sep 17 00:00:00 2001
From: Wenyao Gao <wgao11@u.rochester.edu>
Date: Wed, 22 Jul 2026 23:18:43 -0700
Subject: [PATCH 06/13] refactor(overview): drop unread bucket exactCount

---
 packages/app/src/lib/overview-data.ts | 2 --
 1 file changed, 2 deletions(-)

diff --git a/packages/app/src/lib/overview-data.ts b/packages/app/src/lib/overview-data.ts
index 1e1f9563..8155d686 100644
--- a/packages/app/src/lib/overview-data.ts
+++ b/packages/app/src/lib/overview-data.ts
@@ -272,7 +272,6 @@ interface OverviewHeadlinePairBucket {
   dbModel: string;
   configs: OverviewConfigResult[];
   tierReads: Map<string, ConfigTierRead>;
-  exactCount: number;
   newestEvidence: string;
 }
 
@@ -299,7 +298,6 @@ function buildHeadlinePairBuckets(
         dbModel,
         configs,
         tierReads,
-        exactCount: exactReads.length,
         newestEvidence: exactReads.reduce(
           (latest, read) =>
             read.evidenceDate !== null && read.evidenceDate.to > latest

From 4f38f6b50b2ea46795eabf898d0360f5a7a7e2e3 Mon Sep 17 00:00:00 2001
From: Wenyao Gao <wgao11@u.rochester.edu>
Date: Wed, 22 Jul 2026 23:29:27 -0700
Subject: [PATCH 07/13] fix(header): keep query state when switching language

---
 packages/app/cypress/e2e/overview.cy.ts       |  7 +++++++
 packages/app/src/components/header/header.tsx | 16 +++++++++++++---
 2 files changed, 20 insertions(+), 3 deletions(-)

diff --git a/packages/app/cypress/e2e/overview.cy.ts b/packages/app/cypress/e2e/overview.cy.ts
index 89fb5039..fc63f3be 100644
--- a/packages/app/cypress/e2e/overview.cy.ts
+++ b/packages/app/cypress/e2e/overview.cy.ts
@@ -251,6 +251,13 @@ describe('Overview page', () => {
     cy.get('body')
       .invoke('text')
       .should('not.match', /∞\s*%/);
+
+    cy.visit('/overview?tier=100');
+    cy.contains('Output tok/s/GPU @100 tok/s/user').should('exist');
+    cy.get('[data-testid="language-toggle"]').click();
+    cy.location('pathname').should('eq', '/zh/overview');
+    cy.location('search').should('eq', '?tier=100');
+    cy.contains('每 GPU 输出 tok/s @100 tok/s/用户').should('exist');
   });
 
   it('uses the same cell semantics on mobile and fits both 390px and 320px widths', () => {
diff --git a/packages/app/src/components/header/header.tsx b/packages/app/src/components/header/header.tsx
index 09926859..e029e445 100644
--- a/packages/app/src/components/header/header.tsx
+++ b/packages/app/src/components/header/header.tsx
@@ -75,7 +75,13 @@ function isActive(pathname: string, href: string): boolean {
 }
 
 /** EN ↔ 中文 switcher; maps the current page to its sibling in the other language. */
-function LanguageToggle({ pathname }: { pathname: string }) {
+function LanguageToggle({
+  pathname,
+  router,
+}: {
+  pathname: string;
+  router: ReturnType<typeof useRouter>;
+}) {
   const isZh = isZhPathname(pathname);
   const target = switchLocalePath(pathname);
   return (
@@ -84,7 +90,11 @@ function LanguageToggle({ pathname }: { pathname: string }) {
       data-testid="language-toggle"
       hrefLang={isZh ? 'en' : 'zh-CN'}
       className="inline-flex items-center min-h-11 px-2 rounded-md text-sm font-medium text-muted-foreground hover:text-foreground hover:bg-muted transition-colors whitespace-nowrap"
-      onClick={() => track('header_language_toggled', { to: isZh ? 'en' : 'zh' })}
+      onClick={(event) => {
+        track('header_language_toggled', { to: isZh ? 'en' : 'zh' });
+        // The sibling keeps shareable query state (e.g. the overview's ?tier=).
+        navigateInApp(event, router, target + window.location.search);
+      }}
     >
       {isZh ? 'EN' : '中文'}
     </Link>
@@ -191,7 +201,7 @@ export const Header = ({ starCount }: { starCount?: number | null }) => {
             <span className="hidden sm:flex">
               <GitHubStars owner="SemiAnalysisAI" repo="InferenceX" starCount={starCount} />
             </span>
-            <LanguageToggle pathname={pathname} />
+            <LanguageToggle pathname={pathname} router={router} />
             {/* Minecraft-only audio toggles ride alongside GitHub stars below the
                 same `sm` cutoff: in minecraft mode the pixel font widens the bar,
                 and these two extra buttons are what push a 320px header past its

From 729987537d2074b82f808a800768b600f97cdbce Mon Sep 17 00:00:00 2001
From: Wenyao Gao <wgao11@u.rochester.edu>
Date: Wed, 22 Jul 2026 23:39:18 -0700
Subject: [PATCH 08/13] fix(overview): cannot-reach reason requires every
 speculative bucket unreachable

---
 packages/app/src/lib/overview-data.test.ts | 37 +++++++++++++++++
 packages/app/src/lib/overview-data.ts      | 46 ++++++++++++++--------
 2 files changed, 67 insertions(+), 16 deletions(-)

diff --git a/packages/app/src/lib/overview-data.test.ts b/packages/app/src/lib/overview-data.test.ts
index 29c49a45..61ba6c7f 100644
--- a/packages/app/src/lib/overview-data.test.ts
+++ b/packages/app/src/lib/overview-data.test.ts
@@ -294,6 +294,43 @@ describe('overview headline pairs', () => {
     expect(pair?.highLeaderTransition).toBe('changed_hardware');
   });
 
+  it('claims cannot-reach only when every speculative bucket is unreachable', () => {
+    const unreachable: [number, number][] = [
+      [20, 500],
+      [30, 450],
+      [40, 400],
+      [45, 350],
+    ];
+    const underSwept: [number, number][] = [
+      [60, 900],
+      [70, 800],
+      [80, 700],
+      [90, 600],
+    ];
+    const baseline = frontier([1200, 1000, 800, 600], {
+      hardware: 'b200',
+      precision: Precision.FP4,
+    });
+
+    const mixed = buildOverviewModelSummary(Model.Qwen3_5, [
+      ...baseline,
+      ...frontierAt(unreachable, { hardware: 'mi355x', precision: Precision.FP4 }),
+      ...frontierAt(underSwept, { hardware: 'mi355x', precision: Precision.FP8 }),
+    ]);
+    expect(headlinePairOf(mixed, 'mi355x-vs-b200')?.candidate.missingReason).toBe(
+      'no_exact_at_tier',
+    );
+
+    const allUnreachable = buildOverviewModelSummary(Model.Qwen3_5, [
+      ...baseline,
+      ...frontierAt(unreachable, { hardware: 'mi355x', precision: Precision.FP4 }),
+      ...frontierAt(unreachable, { hardware: 'mi355x', precision: Precision.FP8 }),
+    ]);
+    expect(headlinePairOf(allUnreachable, 'mi355x-vs-b200')?.candidate.missingReason).toBe(
+      'cannot_reach_at_tier',
+    );
+  });
+
   it('distinguishes standard-decode-only and unsupported-precision coverage per member', () => {
     const summary = buildOverviewModelSummary(Model.Qwen3_5, [
       ...frontier([1200, 1000, 800, 600], { hardware: 'b200', precision: Precision.FP4 }),
diff --git a/packages/app/src/lib/overview-data.ts b/packages/app/src/lib/overview-data.ts
index 8155d686..82398216 100644
--- a/packages/app/src/lib/overview-data.ts
+++ b/packages/app/src/lib/overview-data.ts
@@ -318,6 +318,7 @@ function missingReasonForHeadlineMember(
   workloadRows: readonly BenchmarkRow[],
   hardware: string,
   read: OverviewTierRead,
+  bucketReads: readonly OverviewTierRead[],
 ): OverviewMissingReason | null {
   if (isExactTierRead(read)) return null;
   const hardwareRows = workloadRows.filter((row) => row.hardware === hardware);
@@ -325,7 +326,12 @@ function missingReasonForHeadlineMember(
   const supportedRows = hardwareRows.filter((row) => OVERVIEW_PRECISIONS.includes(row.precision));
   if (supportedRows.length === 0) return 'int4_bf16_only';
   if (!supportedRows.some((row) => row.spec_method !== 'none')) return 'standard_decode_only';
-  return read.boundary === 'unreachable' ? 'cannot_reach_at_tier' : 'no_exact_at_tier';
+  // `cannot reach` is a claim about the whole platform, so it holds only when
+  // EVERY qualified speculative stack tops out below the tier — one merely
+  // under-swept stack downgrades the gap to a missing exact read.
+  return bucketReads.length > 0 && bucketReads.every((r) => r.boundary === 'unreachable')
+    ? 'cannot_reach_at_tier'
+    : 'no_exact_at_tier';
 }
 
 function headlineLeaderTransition(
@@ -365,20 +371,23 @@ function buildHeadlinePairs(
   // lexical dbModel. One bucket per member keeps point releases from blending
   // within a read.
   const buildMember = (memberHardware: string): OverviewHeadlinePairMember => {
-    const best = buildHeadlinePairBuckets(configsByPrecision, new Set([memberHardware]), tier)
-      .map((memberBucket) => ({
-        bucket: memberBucket,
-        read: nonComparableAsMissing(memberBucket.tierReads.get(memberHardware), tier),
-      }))
-      .toSorted(
-        (a, b) =>
-          Number(isExactTierRead(b.read)) - Number(isExactTierRead(a.read)) ||
-          (b.read.value ?? -1) - (a.read.value ?? -1) ||
-          OVERVIEW_PRECISIONS.indexOf(a.bucket.precision) -
-            OVERVIEW_PRECISIONS.indexOf(b.bucket.precision) ||
-          b.bucket.newestEvidence.localeCompare(a.bucket.newestEvidence) ||
-          a.bucket.dbModel.localeCompare(b.bucket.dbModel),
-      )[0];
+    const candidates = buildHeadlinePairBuckets(
+      configsByPrecision,
+      new Set([memberHardware]),
+      tier,
+    ).map((memberBucket) => ({
+      bucket: memberBucket,
+      read: nonComparableAsMissing(memberBucket.tierReads.get(memberHardware), tier),
+    }));
+    const best = candidates.toSorted(
+      (a, b) =>
+        Number(isExactTierRead(b.read)) - Number(isExactTierRead(a.read)) ||
+        (b.read.value ?? -1) - (a.read.value ?? -1) ||
+        OVERVIEW_PRECISIONS.indexOf(a.bucket.precision) -
+          OVERVIEW_PRECISIONS.indexOf(b.bucket.precision) ||
+        b.bucket.newestEvidence.localeCompare(a.bucket.newestEvidence) ||
+        a.bucket.dbModel.localeCompare(b.bucket.dbModel),
+    )[0];
     const primary = best?.read ?? nullTierRead(tier);
     const high = nonComparableAsMissing(
       best === undefined
@@ -393,7 +402,12 @@ function buildHeadlinePairs(
       dbModel: best?.bucket.dbModel ?? null,
       read: primary,
       highRead: high,
-      missingReason: missingReasonForHeadlineMember(workloadRows, memberHardware, primary),
+      missingReason: missingReasonForHeadlineMember(
+        workloadRows,
+        memberHardware,
+        primary,
+        candidates.map(({ read }) => read),
+      ),
     };
   };
 

From e2b73249e80a2b04c868e02a25762b846d45bfe9 Mon Sep 17 00:00:00 2001
From: Wenyao Gao <wgao11@u.rochester.edu>
Date: Thu, 23 Jul 2026 10:45:49 -0700
Subject: [PATCH 09/13] feat(overview): mount as first dashboard tab

---
 packages/app/cypress/component/tab-nav.cy.tsx |  6 ++++
 packages/app/cypress/e2e/navigation.cy.ts     |  3 ++
 packages/app/cypress/e2e/overview.cy.ts       | 10 +++++++
 .../app/{ => (dashboard)}/overview/page.tsx   |  8 +----
 .../zh/{ => (dashboard)}/overview/page.tsx    |  8 +----
 packages/app/src/components/header/header.tsx |  1 +
 .../src/components/overview/overview-page.tsx | 30 +++++++++----------
 packages/app/src/components/tab-nav.tsx       |  1 +
 packages/app/src/lib/tab-meta-zh.ts           |  1 +
 9 files changed, 39 insertions(+), 29 deletions(-)
 rename packages/app/src/app/{ => (dashboard)}/overview/page.tsx (87%)
 rename packages/app/src/app/zh/{ => (dashboard)}/overview/page.tsx (88%)

diff --git a/packages/app/cypress/component/tab-nav.cy.tsx b/packages/app/cypress/component/tab-nav.cy.tsx
index 2c24d256..df297e61 100644
--- a/packages/app/cypress/component/tab-nav.cy.tsx
+++ b/packages/app/cypress/component/tab-nav.cy.tsx
@@ -48,6 +48,7 @@ describe('TabNav — unofficialrun URL preservation (issue #319)', () => {
 
   it('renders bare hrefs when the URL has no unofficialrun param', () => {
     mountTabNav({});
+    cy.get('[data-testid="tab-trigger-overview"]').should('have.attr', 'href', '/overview');
     cy.get('[data-testid="tab-trigger-evaluation"]').should('have.attr', 'href', '/evaluation');
     cy.get('[data-testid="tab-trigger-historical"]').should('have.attr', 'href', '/historical');
     cy.get('[data-testid="tab-trigger-calculator"]').should('have.attr', 'href', '/calculator');
@@ -55,6 +56,11 @@ describe('TabNav — unofficialrun URL preservation (issue #319)', () => {
 
   it('appends unofficialruns to every tab href when the URL has the param', () => {
     mountTabNav({ search: '?unofficialruns=12345' });
+    cy.get('[data-testid="tab-trigger-overview"]').should(
+      'have.attr',
+      'href',
+      '/overview?unofficialruns=12345',
+    );
     cy.get('[data-testid="tab-trigger-evaluation"]').should(
       'have.attr',
       'href',
diff --git a/packages/app/cypress/e2e/navigation.cy.ts b/packages/app/cypress/e2e/navigation.cy.ts
index c313c81b..47be8122 100644
--- a/packages/app/cypress/e2e/navigation.cy.ts
+++ b/packages/app/cypress/e2e/navigation.cy.ts
@@ -10,6 +10,9 @@ describe('Chart Section Tabs — E2E', () => {
   });
 
   it('updates the URL path when switching tabs', () => {
+    cy.get('[data-testid="tab-trigger-overview"]').click();
+    cy.url().should('include', '/overview');
+
     cy.get('[data-testid="tab-trigger-evaluation"]').click();
     cy.url().should('include', '/evaluation');
 
diff --git a/packages/app/cypress/e2e/overview.cy.ts b/packages/app/cypress/e2e/overview.cy.ts
index fc63f3be..2219a98e 100644
--- a/packages/app/cypress/e2e/overview.cy.ts
+++ b/packages/app/cypress/e2e/overview.cy.ts
@@ -53,6 +53,12 @@ describe('Overview page', () => {
     cy.viewport(1280, 900);
     cy.visit('/overview');
 
+    cy.get('[data-testid="chart-section-tabs"]').should('be.visible');
+    cy.get('[data-testid="tab-trigger-overview"]')
+      .should('have.attr', 'href', '/overview')
+      .and('have.class', 'border-secondary');
+    cy.get('[data-testid="nav-link-dashboard"]').should('have.class', 'text-brand');
+
     cy.contains('h1', 'AI Inference Overview').should('exist');
     cy.contains(
       'Every active model across MI355X, B200, B300, GB200 and GB300 at a glance.',
@@ -265,6 +271,7 @@ describe('Overview page', () => {
       cy.viewport(width, 844);
       cy.visit('/overview');
 
+      cy.get('[data-testid="mobile-chart-select"]').should('be.visible');
       cy.get('[data-testid="overview-mobile-list"]').should('be.visible');
       cy.get('[data-testid="overview-tier-switcher"]').should('be.visible');
       cy.get('[data-testid="overview-desktop-matrix"]').should('not.be.visible');
@@ -287,6 +294,9 @@ describe('Overview page', () => {
     cy.viewport(1280, 900);
     cy.visit('/zh/overview');
 
+    cy.get('[data-testid="tab-trigger-overview"]')
+      .should('have.attr', 'href', '/zh/overview')
+      .and('contain.text', '总览');
     cy.contains('h1', 'AI 推理总览').should('exist');
     cy.contains('一眼对比各活跃模型在 MI355X、B200、B300、GB200 与 GB300 上的表现。').should(
       'exist',
diff --git a/packages/app/src/app/overview/page.tsx b/packages/app/src/app/(dashboard)/overview/page.tsx
similarity index 87%
rename from packages/app/src/app/overview/page.tsx
rename to packages/app/src/app/(dashboard)/overview/page.tsx
index 8e1bde73..d1c9f8da 100644
--- a/packages/app/src/app/overview/page.tsx
+++ b/packages/app/src/app/(dashboard)/overview/page.tsx
@@ -36,11 +36,5 @@ interface Props {
 export default async function OverviewPage({ searchParams }: Props) {
   const sp = await searchParams;
   const data = await getOverviewPageData(resolveOverviewTier(sp.tier));
-  return (
-    <main className="relative">
-      <div className="container mx-auto px-4 lg:px-8 pb-8">
-        <OverviewPageContent data={data} locale="en" />
-      </div>
-    </main>
-  );
+  return <OverviewPageContent data={data} locale="en" />;
 }
diff --git a/packages/app/src/app/zh/overview/page.tsx b/packages/app/src/app/zh/(dashboard)/overview/page.tsx
similarity index 88%
rename from packages/app/src/app/zh/overview/page.tsx
rename to packages/app/src/app/zh/(dashboard)/overview/page.tsx
index dd5d6cc9..cd94c065 100644
--- a/packages/app/src/app/zh/overview/page.tsx
+++ b/packages/app/src/app/zh/(dashboard)/overview/page.tsx
@@ -37,11 +37,5 @@ interface Props {
 export default async function ZhOverviewPage({ searchParams }: Props) {
   const sp = await searchParams;
   const data = await getOverviewPageData(resolveOverviewTier(sp.tier));
-  return (
-    <main className="relative">
-      <div className="container mx-auto px-4 lg:px-8 pb-8">
-        <OverviewPageContent data={data} locale="zh" />
-      </div>
-    </main>
-  );
+  return <OverviewPageContent data={data} locale="zh" />;
 }
diff --git a/packages/app/src/components/header/header.tsx b/packages/app/src/components/header/header.tsx
index e029e445..02877ffd 100644
--- a/packages/app/src/components/header/header.tsx
+++ b/packages/app/src/components/header/header.tsx
@@ -17,6 +17,7 @@ import { GitHubStars } from './GithubStars';
 
 /** Dashboard tab paths that should highlight the "Dashboard" nav link. */
 const DASHBOARD_TABS = [
+  '/overview',
   '/inference',
   '/evaluation',
   '/historical',
diff --git a/packages/app/src/components/overview/overview-page.tsx b/packages/app/src/components/overview/overview-page.tsx
index 5ac8c864..d9171952 100644
--- a/packages/app/src/components/overview/overview-page.tsx
+++ b/packages/app/src/components/overview/overview-page.tsx
@@ -25,23 +25,23 @@ export function OverviewPageContent({ data, locale }: OverviewPageProps) {
       : strings.snapshot(formatters.shortDate(data.datasetThroughDate));
 
   return (
-    <section className="pt-2 lg:pt-4">
-      <header className="max-w-4xl">
-        <h1 className="text-2xl lg:text-4xl font-bold tracking-tight">{strings.title}</h1>
-        <p className="mt-2 max-w-3xl text-sm text-muted-foreground sm:text-base lg:text-lg">
-          {strings.purpose}
-        </p>
-        <p className="mt-1.5 max-w-4xl text-xs leading-snug text-muted-foreground sm:text-sm">
-          {strings.scope(data.tier)}
-        </p>
-        {snapshot === null ? null : (
-          <p className="mt-1 text-xs text-muted-foreground/80 tabular-nums">{snapshot}</p>
-        )}
-        <OverviewTierSwitcher tier={data.tier} locale={locale} strings={strings} />
-      </header>
+    <section className="flex flex-col gap-4">
+      <Card>
+        <header className="max-w-4xl">
+          <h1 className="text-lg font-semibold mb-2">{strings.title}</h1>
+          <p className="max-w-3xl text-sm text-muted-foreground">{strings.purpose}</p>
+          <p className="mt-1.5 max-w-4xl text-xs leading-snug text-muted-foreground">
+            {strings.scope(data.tier)}
+          </p>
+          {snapshot === null ? null : (
+            <p className="mt-1 text-xs text-muted-foreground/80 tabular-nums">{snapshot}</p>
+          )}
+          <OverviewTierSwitcher tier={data.tier} locale={locale} strings={strings} />
+        </header>
+      </Card>
 
       {/* Official-only summary; uploaded runs remain in the linked dashboard. */}
-      <Card className="mt-4 overflow-hidden p-0 md:p-0">
+      <Card className="overflow-hidden p-0 md:p-0">
         <DesktopOverviewMatrix
           models={data.models}
           locale={locale}
diff --git a/packages/app/src/components/tab-nav.tsx b/packages/app/src/components/tab-nav.tsx
index 0f8d373e..ec99db5f 100644
--- a/packages/app/src/components/tab-nav.tsx
+++ b/packages/app/src/components/tab-nav.tsx
@@ -26,6 +26,7 @@ import { TAB_LABELS_ZH } from '@/lib/tab-meta-zh';
 import { cn } from '@/lib/utils';
 
 const VISIBLE_TABS = [
+  { href: '/overview', label: 'Overview', testId: 'tab-trigger-overview' },
   { href: '/inference', label: 'Inference Performance', testId: 'tab-trigger-inference' },
   { href: '/evaluation', label: 'Accuracy Evals', testId: 'tab-trigger-evaluation' },
   { href: '/historical', label: 'Historical Trends', testId: 'tab-trigger-historical' },
diff --git a/packages/app/src/lib/tab-meta-zh.ts b/packages/app/src/lib/tab-meta-zh.ts
index 7daec602..3c50362d 100644
--- a/packages/app/src/lib/tab-meta-zh.ts
+++ b/packages/app/src/lib/tab-meta-zh.ts
@@ -114,6 +114,7 @@ export const TAB_INTRO_ZH: Record<ZhTabKey, string> = {
 
 /** Chinese labels for the dashboard tab bar (TabNav) on /zh pages. */
 export const TAB_LABELS_ZH: Record<string, string> = {
+  overview: '总览',
   inference: '推理性能',
   evaluation: '准确率评估',
   historical: '历史趋势',

From f3d5ae08206849090eacd105e3f2ea2fd706ad22 Mon Sep 17 00:00:00 2001
From: Wenyao Gao <wgao11@u.rochester.edu>
Date: Thu, 23 Jul 2026 11:02:08 -0700
Subject: [PATCH 10/13] chore(overview): format fixture like other api fixtures

---
 .prettierignore                               |   3 -
 .../cypress/fixtures/api/overview-rows.json   | 630 +++++++++++++++++-
 2 files changed, 609 insertions(+), 24 deletions(-)
 delete mode 100644 .prettierignore

diff --git a/.prettierignore b/.prettierignore
deleted file mode 100644
index 3c258190..00000000
--- a/.prettierignore
+++ /dev/null
@@ -1,3 +0,0 @@
-# Deliberately compact test fixture (one benchmark row per line, <=60 lines);
-# oxfmt/prettier would pretty-print it to ~530 lines. See overview-data.test.ts drift guard.
-packages/app/cypress/fixtures/api/overview-rows.json
diff --git a/packages/app/cypress/fixtures/api/overview-rows.json b/packages/app/cypress/fixtures/api/overview-rows.json
index 96a441b0..feb88fe7 100644
--- a/packages/app/cypress/fixtures/api/overview-rows.json
+++ b/packages/app/cypress/fixtures/api/overview-rows.json
@@ -1,29 +1,617 @@
 {
   "DeepSeek-V4-Pro": [
-    {"id": 1, "hardware": "b300", "framework": "sglang", "model": "dsv4", "precision": "fp4", "spec_method": "mtp", "disagg": false, "is_multinode": false, "prefill_tp": 8, "prefill_ep": 1, "prefill_dp_attention": false, "prefill_num_workers": 1, "decode_tp": 8, "decode_ep": 1, "decode_dp_attention": false, "decode_num_workers": 1, "num_prefill_gpu": 8, "num_decode_gpu": 8, "benchmark_type": "single_turn", "isl": 8192, "osl": 1024, "conc": 8, "offload_mode": "off", "image": null, "metrics": {"median_intvty": 30, "output_tput_per_gpu": 1300}, "date": "2026-06-24", "run_url": null},
-    {"id": 2, "hardware": "b300", "framework": "sglang", "model": "dsv4", "precision": "fp4", "spec_method": "mtp", "disagg": false, "is_multinode": false, "prefill_tp": 8, "prefill_ep": 1, "prefill_dp_attention": false, "prefill_num_workers": 1, "decode_tp": 8, "decode_ep": 1, "decode_dp_attention": false, "decode_num_workers": 1, "num_prefill_gpu": 8, "num_decode_gpu": 8, "benchmark_type": "single_turn", "isl": 8192, "osl": 1024, "conc": 16, "offload_mode": "off", "image": null, "metrics": {"median_intvty": 70, "output_tput_per_gpu": 1000}, "date": "2026-07-04", "run_url": null},
-    {"id": 3, "hardware": "b200", "framework": "sglang", "model": "dsv4", "precision": "fp4", "spec_method": "mtp", "disagg": false, "is_multinode": false, "prefill_tp": 8, "prefill_ep": 1, "prefill_dp_attention": false, "prefill_num_workers": 1, "decode_tp": 8, "decode_ep": 1, "decode_dp_attention": false, "decode_num_workers": 1, "num_prefill_gpu": 8, "num_decode_gpu": 8, "benchmark_type": "single_turn", "isl": 8192, "osl": 1024, "conc": 8, "offload_mode": "off", "image": null, "metrics": {"median_intvty": 40, "output_tput_per_gpu": 1050}, "date": "2026-07-18", "run_url": null},
-    {"id": 4, "hardware": "b200", "framework": "sglang", "model": "dsv4", "precision": "fp4", "spec_method": "mtp", "disagg": false, "is_multinode": false, "prefill_tp": 8, "prefill_ep": 1, "prefill_dp_attention": false, "prefill_num_workers": 1, "decode_tp": 8, "decode_ep": 1, "decode_dp_attention": false, "decode_num_workers": 1, "num_prefill_gpu": 8, "num_decode_gpu": 8, "benchmark_type": "single_turn", "isl": 8192, "osl": 1024, "conc": 24, "offload_mode": "off", "image": null, "metrics": {"median_intvty": 110, "output_tput_per_gpu": 300}, "date": "2026-07-18", "run_url": null},
-    {"id": 5, "hardware": "mi355x", "framework": "sglang", "model": "dsv4", "precision": "fp4", "spec_method": "mtp", "disagg": false, "is_multinode": false, "prefill_tp": 8, "prefill_ep": 1, "prefill_dp_attention": false, "prefill_num_workers": 1, "decode_tp": 8, "decode_ep": 1, "decode_dp_attention": false, "decode_num_workers": 1, "num_prefill_gpu": 8, "num_decode_gpu": 8, "benchmark_type": "single_turn", "isl": 8192, "osl": 1024, "conc": 16, "offload_mode": "off", "image": null, "metrics": {"median_intvty": 60, "output_tput_per_gpu": 760}, "date": "2026-07-18", "run_url": null},
-    {"id": 6, "hardware": "gb300", "framework": "sglang", "model": "dsv4", "precision": "fp4", "spec_method": "mtp", "disagg": false, "is_multinode": false, "prefill_tp": 8, "prefill_ep": 1, "prefill_dp_attention": false, "prefill_num_workers": 1, "decode_tp": 8, "decode_ep": 1, "decode_dp_attention": false, "decode_num_workers": 1, "num_prefill_gpu": 8, "num_decode_gpu": 8, "benchmark_type": "single_turn", "isl": 8192, "osl": 1024, "conc": 8, "offload_mode": "off", "image": null, "metrics": {"median_intvty": 45, "output_tput_per_gpu": 1050}, "date": "2026-07-18", "run_url": null},
-    {"id": 7, "hardware": "h100", "framework": "sglang", "model": "dsv4", "precision": "fp4", "spec_method": "none", "disagg": false, "is_multinode": false, "prefill_tp": 8, "prefill_ep": 1, "prefill_dp_attention": false, "prefill_num_workers": 1, "decode_tp": 8, "decode_ep": 1, "decode_dp_attention": false, "decode_num_workers": 1, "num_prefill_gpu": 8, "num_decode_gpu": 8, "benchmark_type": "single_turn", "isl": 8192, "osl": 1024, "conc": 8, "offload_mode": "off", "image": null, "metrics": {"median_intvty": 30, "output_tput_per_gpu": 800}, "date": "2026-07-18", "run_url": null},
-    {"id": 8, "hardware": "gb200", "framework": "sglang", "model": "dsv4", "precision": "fp8", "spec_method": "mtp", "disagg": false, "is_multinode": false, "prefill_tp": 8, "prefill_ep": 1, "prefill_dp_attention": false, "prefill_num_workers": 1, "decode_tp": 8, "decode_ep": 1, "decode_dp_attention": false, "decode_num_workers": 1, "num_prefill_gpu": 8, "num_decode_gpu": 8, "benchmark_type": "single_turn", "isl": 8192, "osl": 1024, "conc": 8, "offload_mode": "off", "image": null, "metrics": {"median_intvty": 50, "output_tput_per_gpu": 600}, "date": "2026-07-18", "run_url": null}
+    {
+      "id": 1,
+      "hardware": "b300",
+      "framework": "sglang",
+      "model": "dsv4",
+      "precision": "fp4",
+      "spec_method": "mtp",
+      "disagg": false,
+      "is_multinode": false,
+      "prefill_tp": 8,
+      "prefill_ep": 1,
+      "prefill_dp_attention": false,
+      "prefill_num_workers": 1,
+      "decode_tp": 8,
+      "decode_ep": 1,
+      "decode_dp_attention": false,
+      "decode_num_workers": 1,
+      "num_prefill_gpu": 8,
+      "num_decode_gpu": 8,
+      "benchmark_type": "single_turn",
+      "isl": 8192,
+      "osl": 1024,
+      "conc": 8,
+      "offload_mode": "off",
+      "image": null,
+      "metrics": { "median_intvty": 30, "output_tput_per_gpu": 1300 },
+      "date": "2026-06-24",
+      "run_url": null
+    },
+    {
+      "id": 2,
+      "hardware": "b300",
+      "framework": "sglang",
+      "model": "dsv4",
+      "precision": "fp4",
+      "spec_method": "mtp",
+      "disagg": false,
+      "is_multinode": false,
+      "prefill_tp": 8,
+      "prefill_ep": 1,
+      "prefill_dp_attention": false,
+      "prefill_num_workers": 1,
+      "decode_tp": 8,
+      "decode_ep": 1,
+      "decode_dp_attention": false,
+      "decode_num_workers": 1,
+      "num_prefill_gpu": 8,
+      "num_decode_gpu": 8,
+      "benchmark_type": "single_turn",
+      "isl": 8192,
+      "osl": 1024,
+      "conc": 16,
+      "offload_mode": "off",
+      "image": null,
+      "metrics": { "median_intvty": 70, "output_tput_per_gpu": 1000 },
+      "date": "2026-07-04",
+      "run_url": null
+    },
+    {
+      "id": 3,
+      "hardware": "b200",
+      "framework": "sglang",
+      "model": "dsv4",
+      "precision": "fp4",
+      "spec_method": "mtp",
+      "disagg": false,
+      "is_multinode": false,
+      "prefill_tp": 8,
+      "prefill_ep": 1,
+      "prefill_dp_attention": false,
+      "prefill_num_workers": 1,
+      "decode_tp": 8,
+      "decode_ep": 1,
+      "decode_dp_attention": false,
+      "decode_num_workers": 1,
+      "num_prefill_gpu": 8,
+      "num_decode_gpu": 8,
+      "benchmark_type": "single_turn",
+      "isl": 8192,
+      "osl": 1024,
+      "conc": 8,
+      "offload_mode": "off",
+      "image": null,
+      "metrics": { "median_intvty": 40, "output_tput_per_gpu": 1050 },
+      "date": "2026-07-18",
+      "run_url": null
+    },
+    {
+      "id": 4,
+      "hardware": "b200",
+      "framework": "sglang",
+      "model": "dsv4",
+      "precision": "fp4",
+      "spec_method": "mtp",
+      "disagg": false,
+      "is_multinode": false,
+      "prefill_tp": 8,
+      "prefill_ep": 1,
+      "prefill_dp_attention": false,
+      "prefill_num_workers": 1,
+      "decode_tp": 8,
+      "decode_ep": 1,
+      "decode_dp_attention": false,
+      "decode_num_workers": 1,
+      "num_prefill_gpu": 8,
+      "num_decode_gpu": 8,
+      "benchmark_type": "single_turn",
+      "isl": 8192,
+      "osl": 1024,
+      "conc": 24,
+      "offload_mode": "off",
+      "image": null,
+      "metrics": { "median_intvty": 110, "output_tput_per_gpu": 300 },
+      "date": "2026-07-18",
+      "run_url": null
+    },
+    {
+      "id": 5,
+      "hardware": "mi355x",
+      "framework": "sglang",
+      "model": "dsv4",
+      "precision": "fp4",
+      "spec_method": "mtp",
+      "disagg": false,
+      "is_multinode": false,
+      "prefill_tp": 8,
+      "prefill_ep": 1,
+      "prefill_dp_attention": false,
+      "prefill_num_workers": 1,
+      "decode_tp": 8,
+      "decode_ep": 1,
+      "decode_dp_attention": false,
+      "decode_num_workers": 1,
+      "num_prefill_gpu": 8,
+      "num_decode_gpu": 8,
+      "benchmark_type": "single_turn",
+      "isl": 8192,
+      "osl": 1024,
+      "conc": 16,
+      "offload_mode": "off",
+      "image": null,
+      "metrics": { "median_intvty": 60, "output_tput_per_gpu": 760 },
+      "date": "2026-07-18",
+      "run_url": null
+    },
+    {
+      "id": 6,
+      "hardware": "gb300",
+      "framework": "sglang",
+      "model": "dsv4",
+      "precision": "fp4",
+      "spec_method": "mtp",
+      "disagg": false,
+      "is_multinode": false,
+      "prefill_tp": 8,
+      "prefill_ep": 1,
+      "prefill_dp_attention": false,
+      "prefill_num_workers": 1,
+      "decode_tp": 8,
+      "decode_ep": 1,
+      "decode_dp_attention": false,
+      "decode_num_workers": 1,
+      "num_prefill_gpu": 8,
+      "num_decode_gpu": 8,
+      "benchmark_type": "single_turn",
+      "isl": 8192,
+      "osl": 1024,
+      "conc": 8,
+      "offload_mode": "off",
+      "image": null,
+      "metrics": { "median_intvty": 45, "output_tput_per_gpu": 1050 },
+      "date": "2026-07-18",
+      "run_url": null
+    },
+    {
+      "id": 7,
+      "hardware": "h100",
+      "framework": "sglang",
+      "model": "dsv4",
+      "precision": "fp4",
+      "spec_method": "none",
+      "disagg": false,
+      "is_multinode": false,
+      "prefill_tp": 8,
+      "prefill_ep": 1,
+      "prefill_dp_attention": false,
+      "prefill_num_workers": 1,
+      "decode_tp": 8,
+      "decode_ep": 1,
+      "decode_dp_attention": false,
+      "decode_num_workers": 1,
+      "num_prefill_gpu": 8,
+      "num_decode_gpu": 8,
+      "benchmark_type": "single_turn",
+      "isl": 8192,
+      "osl": 1024,
+      "conc": 8,
+      "offload_mode": "off",
+      "image": null,
+      "metrics": { "median_intvty": 30, "output_tput_per_gpu": 800 },
+      "date": "2026-07-18",
+      "run_url": null
+    },
+    {
+      "id": 8,
+      "hardware": "gb200",
+      "framework": "sglang",
+      "model": "dsv4",
+      "precision": "fp8",
+      "spec_method": "mtp",
+      "disagg": false,
+      "is_multinode": false,
+      "prefill_tp": 8,
+      "prefill_ep": 1,
+      "prefill_dp_attention": false,
+      "prefill_num_workers": 1,
+      "decode_tp": 8,
+      "decode_ep": 1,
+      "decode_dp_attention": false,
+      "decode_num_workers": 1,
+      "num_prefill_gpu": 8,
+      "num_decode_gpu": 8,
+      "benchmark_type": "single_turn",
+      "isl": 8192,
+      "osl": 1024,
+      "conc": 8,
+      "offload_mode": "off",
+      "image": null,
+      "metrics": { "median_intvty": 50, "output_tput_per_gpu": 600 },
+      "date": "2026-07-18",
+      "run_url": null
+    }
   ],
   "Qwen-3.5-397B-A17B": [
-    {"id": 9, "hardware": "b300", "framework": "sglang", "model": "qwen3.5", "precision": "fp4", "spec_method": "mtp", "disagg": false, "is_multinode": false, "prefill_tp": 8, "prefill_ep": 1, "prefill_dp_attention": false, "prefill_num_workers": 1, "decode_tp": 8, "decode_ep": 1, "decode_dp_attention": false, "decode_num_workers": 1, "num_prefill_gpu": 8, "num_decode_gpu": 8, "benchmark_type": "single_turn", "isl": 8192, "osl": 1024, "conc": 8, "offload_mode": "off", "image": null, "metrics": {"median_intvty": 30, "output_tput_per_gpu": 1400}, "date": "2026-07-18", "run_url": null},
-    {"id": 10, "hardware": "b300", "framework": "sglang", "model": "qwen3.5", "precision": "fp4", "spec_method": "mtp", "disagg": false, "is_multinode": false, "prefill_tp": 8, "prefill_ep": 1, "prefill_dp_attention": false, "prefill_num_workers": 1, "decode_tp": 8, "decode_ep": 1, "decode_dp_attention": false, "decode_num_workers": 1, "num_prefill_gpu": 8, "num_decode_gpu": 8, "benchmark_type": "single_turn", "isl": 8192, "osl": 1024, "conc": 16, "offload_mode": "off", "image": null, "metrics": {"median_intvty": 70, "output_tput_per_gpu": 980}, "date": "2026-07-18", "run_url": null},
-    {"id": 11, "hardware": "b200", "framework": "sglang", "model": "qwen3.5", "precision": "fp4", "spec_method": "mtp", "disagg": false, "is_multinode": false, "prefill_tp": 8, "prefill_ep": 1, "prefill_dp_attention": false, "prefill_num_workers": 1, "decode_tp": 8, "decode_ep": 1, "decode_dp_attention": false, "decode_num_workers": 1, "num_prefill_gpu": 8, "num_decode_gpu": 8, "benchmark_type": "single_turn", "isl": 8192, "osl": 1024, "conc": 8, "offload_mode": "off", "image": null, "metrics": {"median_intvty": 40, "output_tput_per_gpu": 800}, "date": "2026-07-18", "run_url": null},
-    {"id": 12, "hardware": "b200", "framework": "sglang", "model": "qwen3.5", "precision": "fp4", "spec_method": "mtp", "disagg": false, "is_multinode": false, "prefill_tp": 8, "prefill_ep": 1, "prefill_dp_attention": false, "prefill_num_workers": 1, "decode_tp": 8, "decode_ep": 1, "decode_dp_attention": false, "decode_num_workers": 1, "num_prefill_gpu": 8, "num_decode_gpu": 8, "benchmark_type": "single_turn", "isl": 8192, "osl": 1024, "conc": 24, "offload_mode": "off", "image": null, "metrics": {"median_intvty": 80, "output_tput_per_gpu": 600}, "date": "2026-07-18", "run_url": null},
-    {"id": 13, "hardware": "b200", "framework": "sglang", "model": "qwen3.5", "precision": "fp8", "spec_method": "mtp", "disagg": false, "is_multinode": false, "prefill_tp": 8, "prefill_ep": 1, "prefill_dp_attention": false, "prefill_num_workers": 1, "decode_tp": 8, "decode_ep": 1, "decode_dp_attention": false, "decode_num_workers": 1, "num_prefill_gpu": 8, "num_decode_gpu": 8, "benchmark_type": "single_turn", "isl": 8192, "osl": 1024, "conc": 12, "offload_mode": "off", "image": null, "metrics": {"median_intvty": 50, "output_tput_per_gpu": 900}, "date": "2026-07-18", "run_url": null},
-    {"id": 14, "hardware": "mi355x", "framework": "sglang", "model": "qwen3.5", "precision": "fp8", "spec_method": "mtp", "disagg": false, "is_multinode": false, "prefill_tp": 8, "prefill_ep": 1, "prefill_dp_attention": false, "prefill_num_workers": 1, "decode_tp": 8, "decode_ep": 1, "decode_dp_attention": false, "decode_num_workers": 1, "num_prefill_gpu": 8, "num_decode_gpu": 8, "benchmark_type": "single_turn", "isl": 8192, "osl": 1024, "conc": 12, "offload_mode": "off", "image": null, "metrics": {"median_intvty": 50, "output_tput_per_gpu": 760}, "date": "2026-07-18", "run_url": null},
-    {"id": 19, "hardware": "b200", "framework": "sglang", "model": "qwen3.5", "precision": "fp8", "spec_method": "mtp", "disagg": false, "is_multinode": false, "prefill_tp": 8, "prefill_ep": 1, "prefill_dp_attention": false, "prefill_num_workers": 1, "decode_tp": 8, "decode_ep": 1, "decode_dp_attention": false, "decode_num_workers": 1, "num_prefill_gpu": 8, "num_decode_gpu": 8, "benchmark_type": "single_turn", "isl": 8192, "osl": 1024, "conc": 24, "offload_mode": "off", "image": null, "metrics": {"median_intvty": 120, "output_tput_per_gpu": 300}, "date": "2026-07-18", "run_url": null},
-    {"id": 20, "hardware": "mi355x", "framework": "sglang", "model": "qwen3.5", "precision": "fp8", "spec_method": "mtp", "disagg": false, "is_multinode": false, "prefill_tp": 8, "prefill_ep": 1, "prefill_dp_attention": false, "prefill_num_workers": 1, "decode_tp": 8, "decode_ep": 1, "decode_dp_attention": false, "decode_num_workers": 1, "num_prefill_gpu": 8, "num_decode_gpu": 8, "benchmark_type": "single_turn", "isl": 8192, "osl": 1024, "conc": 24, "offload_mode": "off", "image": null, "metrics": {"median_intvty": 120, "output_tput_per_gpu": 600}, "date": "2026-07-18", "run_url": null}
+    {
+      "id": 9,
+      "hardware": "b300",
+      "framework": "sglang",
+      "model": "qwen3.5",
+      "precision": "fp4",
+      "spec_method": "mtp",
+      "disagg": false,
+      "is_multinode": false,
+      "prefill_tp": 8,
+      "prefill_ep": 1,
+      "prefill_dp_attention": false,
+      "prefill_num_workers": 1,
+      "decode_tp": 8,
+      "decode_ep": 1,
+      "decode_dp_attention": false,
+      "decode_num_workers": 1,
+      "num_prefill_gpu": 8,
+      "num_decode_gpu": 8,
+      "benchmark_type": "single_turn",
+      "isl": 8192,
+      "osl": 1024,
+      "conc": 8,
+      "offload_mode": "off",
+      "image": null,
+      "metrics": { "median_intvty": 30, "output_tput_per_gpu": 1400 },
+      "date": "2026-07-18",
+      "run_url": null
+    },
+    {
+      "id": 10,
+      "hardware": "b300",
+      "framework": "sglang",
+      "model": "qwen3.5",
+      "precision": "fp4",
+      "spec_method": "mtp",
+      "disagg": false,
+      "is_multinode": false,
+      "prefill_tp": 8,
+      "prefill_ep": 1,
+      "prefill_dp_attention": false,
+      "prefill_num_workers": 1,
+      "decode_tp": 8,
+      "decode_ep": 1,
+      "decode_dp_attention": false,
+      "decode_num_workers": 1,
+      "num_prefill_gpu": 8,
+      "num_decode_gpu": 8,
+      "benchmark_type": "single_turn",
+      "isl": 8192,
+      "osl": 1024,
+      "conc": 16,
+      "offload_mode": "off",
+      "image": null,
+      "metrics": { "median_intvty": 70, "output_tput_per_gpu": 980 },
+      "date": "2026-07-18",
+      "run_url": null
+    },
+    {
+      "id": 11,
+      "hardware": "b200",
+      "framework": "sglang",
+      "model": "qwen3.5",
+      "precision": "fp4",
+      "spec_method": "mtp",
+      "disagg": false,
+      "is_multinode": false,
+      "prefill_tp": 8,
+      "prefill_ep": 1,
+      "prefill_dp_attention": false,
+      "prefill_num_workers": 1,
+      "decode_tp": 8,
+      "decode_ep": 1,
+      "decode_dp_attention": false,
+      "decode_num_workers": 1,
+      "num_prefill_gpu": 8,
+      "num_decode_gpu": 8,
+      "benchmark_type": "single_turn",
+      "isl": 8192,
+      "osl": 1024,
+      "conc": 8,
+      "offload_mode": "off",
+      "image": null,
+      "metrics": { "median_intvty": 40, "output_tput_per_gpu": 800 },
+      "date": "2026-07-18",
+      "run_url": null
+    },
+    {
+      "id": 12,
+      "hardware": "b200",
+      "framework": "sglang",
+      "model": "qwen3.5",
+      "precision": "fp4",
+      "spec_method": "mtp",
+      "disagg": false,
+      "is_multinode": false,
+      "prefill_tp": 8,
+      "prefill_ep": 1,
+      "prefill_dp_attention": false,
+      "prefill_num_workers": 1,
+      "decode_tp": 8,
+      "decode_ep": 1,
+      "decode_dp_attention": false,
+      "decode_num_workers": 1,
+      "num_prefill_gpu": 8,
+      "num_decode_gpu": 8,
+      "benchmark_type": "single_turn",
+      "isl": 8192,
+      "osl": 1024,
+      "conc": 24,
+      "offload_mode": "off",
+      "image": null,
+      "metrics": { "median_intvty": 80, "output_tput_per_gpu": 600 },
+      "date": "2026-07-18",
+      "run_url": null
+    },
+    {
+      "id": 13,
+      "hardware": "b200",
+      "framework": "sglang",
+      "model": "qwen3.5",
+      "precision": "fp8",
+      "spec_method": "mtp",
+      "disagg": false,
+      "is_multinode": false,
+      "prefill_tp": 8,
+      "prefill_ep": 1,
+      "prefill_dp_attention": false,
+      "prefill_num_workers": 1,
+      "decode_tp": 8,
+      "decode_ep": 1,
+      "decode_dp_attention": false,
+      "decode_num_workers": 1,
+      "num_prefill_gpu": 8,
+      "num_decode_gpu": 8,
+      "benchmark_type": "single_turn",
+      "isl": 8192,
+      "osl": 1024,
+      "conc": 12,
+      "offload_mode": "off",
+      "image": null,
+      "metrics": { "median_intvty": 50, "output_tput_per_gpu": 900 },
+      "date": "2026-07-18",
+      "run_url": null
+    },
+    {
+      "id": 14,
+      "hardware": "mi355x",
+      "framework": "sglang",
+      "model": "qwen3.5",
+      "precision": "fp8",
+      "spec_method": "mtp",
+      "disagg": false,
+      "is_multinode": false,
+      "prefill_tp": 8,
+      "prefill_ep": 1,
+      "prefill_dp_attention": false,
+      "prefill_num_workers": 1,
+      "decode_tp": 8,
+      "decode_ep": 1,
+      "decode_dp_attention": false,
+      "decode_num_workers": 1,
+      "num_prefill_gpu": 8,
+      "num_decode_gpu": 8,
+      "benchmark_type": "single_turn",
+      "isl": 8192,
+      "osl": 1024,
+      "conc": 12,
+      "offload_mode": "off",
+      "image": null,
+      "metrics": { "median_intvty": 50, "output_tput_per_gpu": 760 },
+      "date": "2026-07-18",
+      "run_url": null
+    },
+    {
+      "id": 19,
+      "hardware": "b200",
+      "framework": "sglang",
+      "model": "qwen3.5",
+      "precision": "fp8",
+      "spec_method": "mtp",
+      "disagg": false,
+      "is_multinode": false,
+      "prefill_tp": 8,
+      "prefill_ep": 1,
+      "prefill_dp_attention": false,
+      "prefill_num_workers": 1,
+      "decode_tp": 8,
+      "decode_ep": 1,
+      "decode_dp_attention": false,
+      "decode_num_workers": 1,
+      "num_prefill_gpu": 8,
+      "num_decode_gpu": 8,
+      "benchmark_type": "single_turn",
+      "isl": 8192,
+      "osl": 1024,
+      "conc": 24,
+      "offload_mode": "off",
+      "image": null,
+      "metrics": { "median_intvty": 120, "output_tput_per_gpu": 300 },
+      "date": "2026-07-18",
+      "run_url": null
+    },
+    {
+      "id": 20,
+      "hardware": "mi355x",
+      "framework": "sglang",
+      "model": "qwen3.5",
+      "precision": "fp8",
+      "spec_method": "mtp",
+      "disagg": false,
+      "is_multinode": false,
+      "prefill_tp": 8,
+      "prefill_ep": 1,
+      "prefill_dp_attention": false,
+      "prefill_num_workers": 1,
+      "decode_tp": 8,
+      "decode_ep": 1,
+      "decode_dp_attention": false,
+      "decode_num_workers": 1,
+      "num_prefill_gpu": 8,
+      "num_decode_gpu": 8,
+      "benchmark_type": "single_turn",
+      "isl": 8192,
+      "osl": 1024,
+      "conc": 24,
+      "offload_mode": "off",
+      "image": null,
+      "metrics": { "median_intvty": 120, "output_tput_per_gpu": 600 },
+      "date": "2026-07-18",
+      "run_url": null
+    }
   ],
   "MiniMax-M3": [
-    {"id": 15, "hardware": "h200", "framework": "sglang", "model": "minimaxm3", "precision": "fp8", "spec_method": "mtp", "disagg": false, "is_multinode": false, "prefill_tp": 8, "prefill_ep": 1, "prefill_dp_attention": false, "prefill_num_workers": 1, "decode_tp": 8, "decode_ep": 1, "decode_dp_attention": false, "decode_num_workers": 1, "num_prefill_gpu": 8, "num_decode_gpu": 8, "benchmark_type": "single_turn", "isl": 8192, "osl": 1024, "conc": 8, "offload_mode": "off", "image": null, "metrics": {"median_intvty": 30, "output_tput_per_gpu": 820}, "date": "2026-07-18", "run_url": null},
-    {"id": 16, "hardware": "h200", "framework": "sglang", "model": "minimaxm3", "precision": "fp8", "spec_method": "mtp", "disagg": false, "is_multinode": false, "prefill_tp": 8, "prefill_ep": 1, "prefill_dp_attention": false, "prefill_num_workers": 1, "decode_tp": 8, "decode_ep": 1, "decode_dp_attention": false, "decode_num_workers": 1, "num_prefill_gpu": 8, "num_decode_gpu": 8, "benchmark_type": "single_turn", "isl": 8192, "osl": 1024, "conc": 16, "offload_mode": "off", "image": null, "metrics": {"median_intvty": 70, "output_tput_per_gpu": 560}, "date": "2026-07-18", "run_url": null},
-    {"id": 17, "hardware": "mi325x", "framework": "sglang", "model": "minimaxm3", "precision": "fp8", "spec_method": "mtp", "disagg": false, "is_multinode": false, "prefill_tp": 8, "prefill_ep": 1, "prefill_dp_attention": false, "prefill_num_workers": 1, "decode_tp": 8, "decode_ep": 1, "decode_dp_attention": false, "decode_num_workers": 1, "num_prefill_gpu": 8, "num_decode_gpu": 8, "benchmark_type": "single_turn", "isl": 8192, "osl": 1024, "conc": 12, "offload_mode": "off", "image": null, "metrics": {"median_intvty": 50, "output_tput_per_gpu": 650}, "date": "2026-07-18", "run_url": null},
-    {"id": 18, "hardware": "h200", "framework": "sglang", "model": "minimaxm3", "precision": "fp4", "spec_method": "mtp", "disagg": false, "is_multinode": false, "prefill_tp": 8, "prefill_ep": 1, "prefill_dp_attention": false, "prefill_num_workers": 1, "decode_tp": 8, "decode_ep": 1, "decode_dp_attention": false, "decode_num_workers": 1, "num_prefill_gpu": 8, "num_decode_gpu": 8, "benchmark_type": "single_turn", "isl": 8192, "osl": 1024, "conc": 12, "offload_mode": "off", "image": null, "metrics": {"median_intvty": 50, "output_tput_per_gpu": 600}, "date": "2026-07-18", "run_url": null},
-    {"id": 21, "hardware": "gb300", "framework": "sglang", "model": "minimaxm3", "precision": "fp8", "spec_method": "mtp", "disagg": false, "is_multinode": false, "prefill_tp": 8, "prefill_ep": 1, "prefill_dp_attention": false, "prefill_num_workers": 1, "decode_tp": 8, "decode_ep": 1, "decode_dp_attention": false, "decode_num_workers": 1, "num_prefill_gpu": 8, "num_decode_gpu": 8, "benchmark_type": "single_turn", "isl": 8192, "osl": 1024, "conc": 12, "offload_mode": "off", "image": null, "metrics": {"median_intvty": 50, "output_tput_per_gpu": 700}, "date": "2026-07-18", "run_url": null}
+    {
+      "id": 15,
+      "hardware": "h200",
+      "framework": "sglang",
+      "model": "minimaxm3",
+      "precision": "fp8",
+      "spec_method": "mtp",
+      "disagg": false,
+      "is_multinode": false,
+      "prefill_tp": 8,
+      "prefill_ep": 1,
+      "prefill_dp_attention": false,
+      "prefill_num_workers": 1,
+      "decode_tp": 8,
+      "decode_ep": 1,
+      "decode_dp_attention": false,
+      "decode_num_workers": 1,
+      "num_prefill_gpu": 8,
+      "num_decode_gpu": 8,
+      "benchmark_type": "single_turn",
+      "isl": 8192,
+      "osl": 1024,
+      "conc": 8,
+      "offload_mode": "off",
+      "image": null,
+      "metrics": { "median_intvty": 30, "output_tput_per_gpu": 820 },
+      "date": "2026-07-18",
+      "run_url": null
+    },
+    {
+      "id": 16,
+      "hardware": "h200",
+      "framework": "sglang",
+      "model": "minimaxm3",
+      "precision": "fp8",
+      "spec_method": "mtp",
+      "disagg": false,
+      "is_multinode": false,
+      "prefill_tp": 8,
+      "prefill_ep": 1,
+      "prefill_dp_attention": false,
+      "prefill_num_workers": 1,
+      "decode_tp": 8,
+      "decode_ep": 1,
+      "decode_dp_attention": false,
+      "decode_num_workers": 1,
+      "num_prefill_gpu": 8,
+      "num_decode_gpu": 8,
+      "benchmark_type": "single_turn",
+      "isl": 8192,
+      "osl": 1024,
+      "conc": 16,
+      "offload_mode": "off",
+      "image": null,
+      "metrics": { "median_intvty": 70, "output_tput_per_gpu": 560 },
+      "date": "2026-07-18",
+      "run_url": null
+    },
+    {
+      "id": 17,
+      "hardware": "mi325x",
+      "framework": "sglang",
+      "model": "minimaxm3",
+      "precision": "fp8",
+      "spec_method": "mtp",
+      "disagg": false,
+      "is_multinode": false,
+      "prefill_tp": 8,
+      "prefill_ep": 1,
+      "prefill_dp_attention": false,
+      "prefill_num_workers": 1,
+      "decode_tp": 8,
+      "decode_ep": 1,
+      "decode_dp_attention": false,
+      "decode_num_workers": 1,
+      "num_prefill_gpu": 8,
+      "num_decode_gpu": 8,
+      "benchmark_type": "single_turn",
+      "isl": 8192,
+      "osl": 1024,
+      "conc": 12,
+      "offload_mode": "off",
+      "image": null,
+      "metrics": { "median_intvty": 50, "output_tput_per_gpu": 650 },
+      "date": "2026-07-18",
+      "run_url": null
+    },
+    {
+      "id": 18,
+      "hardware": "h200",
+      "framework": "sglang",
+      "model": "minimaxm3",
+      "precision": "fp4",
+      "spec_method": "mtp",
+      "disagg": false,
+      "is_multinode": false,
+      "prefill_tp": 8,
+      "prefill_ep": 1,
+      "prefill_dp_attention": false,
+      "prefill_num_workers": 1,
+      "decode_tp": 8,
+      "decode_ep": 1,
+      "decode_dp_attention": false,
+      "decode_num_workers": 1,
+      "num_prefill_gpu": 8,
+      "num_decode_gpu": 8,
+      "benchmark_type": "single_turn",
+      "isl": 8192,
+      "osl": 1024,
+      "conc": 12,
+      "offload_mode": "off",
+      "image": null,
+      "metrics": { "median_intvty": 50, "output_tput_per_gpu": 600 },
+      "date": "2026-07-18",
+      "run_url": null
+    },
+    {
+      "id": 21,
+      "hardware": "gb300",
+      "framework": "sglang",
+      "model": "minimaxm3",
+      "precision": "fp8",
+      "spec_method": "mtp",
+      "disagg": false,
+      "is_multinode": false,
+      "prefill_tp": 8,
+      "prefill_ep": 1,
+      "prefill_dp_attention": false,
+      "prefill_num_workers": 1,
+      "decode_tp": 8,
+      "decode_ep": 1,
+      "decode_dp_attention": false,
+      "decode_num_workers": 1,
+      "num_prefill_gpu": 8,
+      "num_decode_gpu": 8,
+      "benchmark_type": "single_turn",
+      "isl": 8192,
+      "osl": 1024,
+      "conc": 12,
+      "offload_mode": "off",
+      "image": null,
+      "metrics": { "median_intvty": 50, "output_tput_per_gpu": 700 },
+      "date": "2026-07-18",
+      "run_url": null
+    }
   ]
 }

From 69a0a1a3774d7b74421a8f7e3918f0b58a4f56e3 Mon Sep 17 00:00:00 2001
From: Wenyao Gao <wgao11@u.rochester.edu>
Date: Thu, 23 Jul 2026 11:09:55 -0700
Subject: [PATCH 11/13] chore(overview): trim comments to load-bearing notes

---
 packages/app/cypress/component/header.cy.tsx  |   6 +-
 packages/app/src/components/header/header.tsx |   7 +-
 .../overview/overview-detail-link.tsx         |   9 +-
 .../overview/overview-scorecard.tsx           |  55 +--------
 .../app/src/lib/overview-config-identity.ts   |  12 +-
 packages/app/src/lib/overview-data.server.ts  |   7 +-
 packages/app/src/lib/overview-data.test.ts    |  19 +--
 packages/app/src/lib/overview-data.ts         | 109 ++++--------------
 packages/app/src/lib/overview-links.ts        |   7 --
 9 files changed, 45 insertions(+), 186 deletions(-)

diff --git a/packages/app/cypress/component/header.cy.tsx b/packages/app/cypress/component/header.cy.tsx
index 308b31d4..070fcfa8 100644
--- a/packages/app/cypress/component/header.cy.tsx
+++ b/packages/app/cypress/component/header.cy.tsx
@@ -5,10 +5,8 @@ import { PathnameContext } from 'next/dist/shared/lib/hooks-client-context.share
 import { Header } from '@/components/header/header';
 import { ThemeProvider } from '@/components/ui/theme-provider';
 
-// Component specs mount outside the Next app shell, so the global stylesheet is
-// never injected. The layout assertions below need real Tailwind output, and
-// next-style-loader inserts before this anchor node, so both must exist before
-// the stylesheet module evaluates.
+// Mounted outside the Next app shell; next-style-loader inserts the global
+// stylesheet before this anchor, so it must exist before the import below.
 const cssAnchor = document.createElement('noscript');
 cssAnchor.id = '__next_css__DO_NOT_USE__';
 document.head.append(cssAnchor);
diff --git a/packages/app/src/components/header/header.tsx b/packages/app/src/components/header/header.tsx
index 02877ffd..bc438d34 100644
--- a/packages/app/src/components/header/header.tsx
+++ b/packages/app/src/components/header/header.tsx
@@ -203,11 +203,8 @@ export const Header = ({ starCount }: { starCount?: number | null }) => {
               <GitHubStars owner="SemiAnalysisAI" repo="InferenceX" starCount={starCount} />
             </span>
             <LanguageToggle pathname={pathname} router={router} />
-            {/* Minecraft-only audio toggles ride alongside GitHub stars below the
-                same `sm` cutoff: in minecraft mode the pixel font widens the bar,
-                and these two extra buttons are what push a 320px header past its
-                bounds. Below `sm` they move into the mobile menu instead (see the
-                MinecraftToggles instance in the hamburger dropdown below). */}
+            {/* Below `sm` these move into the mobile menu — they are what push
+                a 320px header past its bounds in minecraft mode. */}
             <span className="hidden items-center gap-2 sm:flex">
               <MinecraftToggles />
             </span>
diff --git a/packages/app/src/components/overview/overview-detail-link.tsx b/packages/app/src/components/overview/overview-detail-link.tsx
index 2e1e44cb..dbd11283 100644
--- a/packages/app/src/components/overview/overview-detail-link.tsx
+++ b/packages/app/src/components/overview/overview-detail-link.tsx
@@ -6,12 +6,9 @@ import type { ReactNode } from 'react';
 import { track } from '@/lib/analytics';
 import { cn } from '@/lib/utils';
 
-/**
- * Model-level dashboard link. A plain anchor, not a `<Link>`: the dashboard
- * reads its filters from a snapshot `url-state.ts` takes at module evaluation,
- * so a client-side navigation would land on an unfiltered dashboard while the
- * label promises the model's own view.
- */
+/** A plain anchor, not `<Link>`: the dashboard reads filters from a snapshot
+ *  `url-state.ts` takes at module evaluation, so client-side navigation would
+ *  land unfiltered. */
 export function OverviewDetailLink({
   href,
   model,
diff --git a/packages/app/src/components/overview/overview-scorecard.tsx b/packages/app/src/components/overview/overview-scorecard.tsx
index a6adb504..20605187 100644
--- a/packages/app/src/components/overview/overview-scorecard.tsx
+++ b/packages/app/src/components/overview/overview-scorecard.tsx
@@ -1,20 +1,3 @@
-/**
- * @file overview-scorecard.tsx
- * @description The comparison surface itself: a portfolio matrix of every
- * active model (rows) across today's key serving platforms (columns — B200 as
- * the reference, then MI355X, B300, GB200 NVL72, GB300 NVL72). The desktop
- * table at `xl` and the mobile cards below `xl` render every cell through the
- * SAME `BaselineCell`/`CandidateCell` components, so the two surfaces cannot
- * drift semantically.
- *
- * Each cell shows that platform's own best validated speculative read at the
- * displayed tier (50 tok/s/user by default; `?tier=` picks another) with its
- * precision badge, evidence date and filtered-dashboard link. The signed
- * delta against B200 appears only when both displayed reads share a precision
- * and a release — mismatches say so instead of comparing FP4 with FP8. A side
- * with no exact read shows `∞` with its reason, never a number or percent.
- */
-
 import {
   OVERVIEW_HIGH_TIER,
   OVERVIEW_TIERS,
@@ -118,11 +101,6 @@ export function overviewFormatters(locale: OverviewLocale): Formatters {
   };
 }
 
-/**
- * A result's evidence dates, localized: a single day (`Jul 6` / `7月6日`) when
- * both backing frontier points share a date, else an en-dash range
- * (`Jun 24–Jul 4`). The caller renders nothing when there is no backing date.
- */
 function formatEvidenceDate(
   formatters: Formatters,
   evidenceDate: { from: string; to: string },
@@ -133,24 +111,18 @@ function formatEvidenceDate(
     : `${from}–${formatters.shortDate(evidenceDate.to)}`;
 }
 
-/** The reason copy for a missing side, keyed to the displayed tier. */
 function missingReasonCopy(member: OverviewHeadlinePairMember, strings: OverviewStrings): string {
   const reason = member.missingReason;
   return reason === null ? '' : strings.missingReasons(member.read.tier)[reason];
 }
 
-/** `candidate/B200 − 1` as a rounded percent with a U+2212 or `+` sign. */
 function signedPercent(formatters: Formatters, percent: number): string {
   const rounded = Math.round(percent);
   const sign = rounded < 0 ? '−' : '+';
   return `${sign}${formatters.number.format(Math.abs(rounded))}%`;
 }
 
-/**
- * The @100 line, rendered only when the @100 leader is a different hardware
- * than the displayed tier's leader — the leader being the side with the
- * greater exact @100 read. Null on the 100 view itself (data layer).
- */
+/** Rendered only when the @100 leader flips; the data layer nulls it on the 100 view. */
 function highLeadLine(
   pair: OverviewHeadlinePairComparison,
   strings: OverviewStrings,
@@ -167,7 +139,6 @@ function highLeadLine(
 const PAIR_VALUE_LINK_CLASS =
   'inline-flex min-h-11 items-center rounded-sm underline decoration-dotted underline-offset-4 hover:decoration-solid focus-visible:outline-none focus-visible:ring-[3px] focus-visible:ring-ring/50';
 
-/** A side with no exact read at the displayed tier: `∞` plus its reason, no link/date/percent. */
 function CellMissing({ hardware, reason }: { hardware: string; reason: string }) {
   return (
     <span
@@ -182,11 +153,6 @@ function CellMissing({ hardware, reason }: { hardware: string; reason: string })
   );
 }
 
-/**
- * One platform's read: its exact displayed-tier value linked to the platform's
- * own filtered dashboard, with a precision badge and the evidence date — or
- * `∞` with the reason it has no exact read.
- */
 function CellValue({
   locale,
   model,
@@ -204,8 +170,7 @@ function CellValue({
   if (member.missingReason !== null || value === null) {
     return <CellMissing hardware={member.hardware} reason={missingReasonCopy(member, strings)} />;
   }
-  // The full deployment stack behind this read; the framework and precision
-  // stay visible, the spec method rides the link title and aria only.
+  // Framework and precision stay visible; the spec method rides the title/aria only.
   const stack =
     config === null
       ? null
@@ -250,7 +215,6 @@ function CellValue({
   );
 }
 
-/** The B200 reference cell: the read alone, no delta against itself. */
 function BaselineCell(props: {
   locale: OverviewLocale;
   model: OverviewModelSummary;
@@ -265,11 +229,6 @@ function BaselineCell(props: {
   );
 }
 
-/**
- * One candidate platform's cell: its own read, then the same-precision delta
- * against B200, a mismatch note when both sides exist but cannot compare, and
- * the @100 leader line only when the leader flips.
- */
 function CandidateCell({
   locale,
   model,
@@ -316,7 +275,6 @@ function CandidateCell({
   );
 }
 
-/** The model's display name, shared by both surfaces. */
 function ModelName({ model }: { model: OverviewModelSummary }) {
   return <h2 className="text-sm font-semibold leading-snug">{model.modelLabel}</h2>;
 }
@@ -328,7 +286,6 @@ interface SurfaceProps {
   strings: OverviewStrings;
 }
 
-/** Model column, the B200 reference, one column per candidate platform, details. */
 export function DesktopOverviewMatrix({ models, locale, formatters, strings }: SurfaceProps) {
   const headlinePairs = models[0]?.headlinePairs ?? [];
   const baselineHeader = headlinePairs[0]?.baseline.hardwareLabel ?? null;
@@ -417,7 +374,6 @@ export function DesktopOverviewMatrix({ models, locale, formatters, strings }: S
   );
 }
 
-/** Below `xl`: the same cells stacked as one card per model, always fully visible. */
 export function MobileOverviewList({ models, locale, formatters, strings }: SurfaceProps) {
   return (
     <ul data-testid="overview-mobile-list" className="divide-y divide-border/50 xl:hidden">
@@ -474,11 +430,8 @@ export function MobileOverviewList({ models, locale, formatters, strings }: Surf
   );
 }
 
-/**
- * The service-level selector: one plain link per tier, so every view is a
- * server-rendered, copyable URL — no client state. The displayed tier is
- * inert text marked `aria-current`, never a self-link.
- */
+/** Plain links so every view is a copyable server-rendered URL; the displayed
+ *  tier is inert `aria-current` text, never a self-link. */
 export function OverviewTierSwitcher({
   tier,
   locale,
diff --git a/packages/app/src/lib/overview-config-identity.ts b/packages/app/src/lib/overview-config-identity.ts
index 18e4e9f4..c5b5948b 100644
--- a/packages/app/src/lib/overview-config-identity.ts
+++ b/packages/app/src/lib/overview-config-identity.ts
@@ -1,14 +1,8 @@
 import type { BenchmarkRow } from './api';
 
-/**
- * Exact per-deployment identity for the overview grouping. Every dimension that
- * makes a serving configuration a distinct deployable topology is part of the
- * key — model, hardware, framework, precision, spec method, disagg, multinode,
- * per-role parallelism and worker counts, GPU counts, and offload mode. Rows
- * that differ only in concurrency, date, image, or run URL share one identity.
- *
- * Serialized as a JSON tuple so no field delimiter can collide with a value.
- */
+/** Every dimension that makes a distinct deployable topology; rows differing
+ *  only in concurrency, date, image, or run URL share one identity. A JSON
+ *  tuple so no delimiter can collide with a value. */
 export function overviewConfigIdentityKey(row: BenchmarkRow): string {
   return JSON.stringify([
     row.model,
diff --git a/packages/app/src/lib/overview-data.server.ts b/packages/app/src/lib/overview-data.server.ts
index 0bfa5fe6..a7fc133c 100644
--- a/packages/app/src/lib/overview-data.server.ts
+++ b/packages/app/src/lib/overview-data.server.ts
@@ -15,9 +15,8 @@ import { loadFixture } from '@/lib/test-fixtures';
 export async function getOverviewPageData(
   tier: OverviewTier = OVERVIEW_PRIMARY_TIER,
 ): Promise<OverviewPageData> {
-  // E2E fixtures mode serves a small synthetic rows-by-display-model fixture
-  // through the same assembler the live path uses, so a contract drift breaks
-  // the fixture tests instead of silently stranding the page's data.
+  // Synthetic rows go through the same assembler as the live path, so a
+  // contract drift breaks the fixture tests instead of stranding the page.
   if (FIXTURES_MODE) {
     return assembleOverviewPageData(
       loadFixture<Record<string, BenchmarkRow[]>>('overview-rows'),
@@ -25,8 +24,6 @@ export async function getOverviewPageData(
     );
   }
 
-  // Fetch rows per db model, concatenated per display model (one display model
-  // can span several db buckets), then hand the same shape to the assembler.
   const entries = await Promise.all(
     [...DEFAULT_MODELS].map(async (model) => {
       const keys = DISPLAY_MODEL_TO_DB[model] ?? [];
diff --git a/packages/app/src/lib/overview-data.test.ts b/packages/app/src/lib/overview-data.test.ts
index 61ba6c7f..da4f03d7 100644
--- a/packages/app/src/lib/overview-data.test.ts
+++ b/packages/app/src/lib/overview-data.test.ts
@@ -482,11 +482,8 @@ describe('tier-parameterized overview', () => {
   });
 });
 
-// Drift guard for the synthetic overview e2e fixture: if the assembler contract
-// changes, the hand-built rows in overview-rows.json must be regenerated, and
-// this fails loudly instead of the e2e stranding empty data. Every expectation
-// is derived by running this same assembler over the fixture — never eyeballed —
-// so it locks the exact matrix cell states overview.cy.ts renders against.
+// Drift guard: runs the real assembler over the e2e fixture; expectations are
+// engine-derived, never eyeballed. Contract drift fails here, not in overview.cy.ts.
 describe('assembleOverviewPageData over the overview-rows fixture', () => {
   it('serves every matrix cell state through the real builder', () => {
     const page = assembleOverviewPageData(
@@ -497,10 +494,8 @@ describe('assembleOverviewPageData over the overview-rows fixture', () => {
     expect(page.datasetThroughDate).toBe('2026-07-18');
     expect(page.tier).toBe(50);
 
-    // DeepSeek: FP4 exacts on B200/B300 carry the only same-precision delta with
-    // a cross-day evidence range; GB200's own best is FP8, so its value shows
-    // but the delta is withheld; MI355X and GB300 miss in opposite clamp
-    // directions.
+    // DeepSeek: FP4 delta + cross-day range; GB200's FP8 best withholds its
+    // delta; MI355X and GB300 miss in opposite clamp directions.
     const deepseek = page.models.find((m) => m.model === Model.DeepSeek_V4_Pro)!;
     const dsB300 = headlinePairOf(deepseek, 'b300-vs-b200')!;
     expect(dsB300.precision).toBe(Precision.FP4);
@@ -519,8 +514,7 @@ describe('assembleOverviewPageData over the overview-rows fixture', () => {
       'cannot_reach_at_tier',
     );
 
-    // MiniMax: only GB300 has workload rows, so its value stands alone — a
-    // missing baseline never yields a delta OR a mismatch note.
+    // MiniMax: a missing baseline yields neither delta nor mismatch note.
     const minimax = page.models.find((m) => m.model === Model.MiniMax_M3)!;
     const mmGb300 = headlinePairOf(minimax, 'gb300-vs-b200')!;
     expect(mmGb300.baseline.missingReason).toBe('no_8k1k_data');
@@ -528,8 +522,7 @@ describe('assembleOverviewPageData over the overview-rows fixture', () => {
     expect(mmGb300.directDeltaPercent).toBeNull();
     expect(mmGb300.deltaUnavailableReason).toBeNull();
 
-    // Qwen: FP8 exacts on both sides yield the −16% delta and a leader flip at
-    // 100, while B300's FP4 best withholds its delta against the FP8 baseline.
+    // Qwen: FP8 delta −16% + leader flip at 100; B300's FP4 best withholds its delta.
     const qwen = page.models.find((m) => m.model === Model.Qwen3_5)!;
     const qwenMi = headlinePairOf(qwen, 'mi355x-vs-b200')!;
     expect(qwenMi.precision).toBe(Precision.FP8);
diff --git a/packages/app/src/lib/overview-data.ts b/packages/app/src/lib/overview-data.ts
index 82398216..f74d3fb0 100644
--- a/packages/app/src/lib/overview-data.ts
+++ b/packages/app/src/lib/overview-data.ts
@@ -10,12 +10,9 @@ import { computeTcoFeed, type TcoTierBoundary } from './tco-feed';
 export const OVERVIEW_WORKLOAD = { isl: 8192, osl: 1024 } as const;
 export const OVERVIEW_TIERS = [30, 50, 75, 100] as const;
 export type OverviewTier = (typeof OVERVIEW_TIERS)[number];
-/** Default service point; ?tier= re-renders the matrix at any OVERVIEW_TIERS value. */
 export const OVERVIEW_PRIMARY_TIER = 50;
-/** High-interactivity capability read; also drives the high-tier leader transition. */
 export const OVERVIEW_HIGH_TIER = 100;
 
-/** The `?tier=` query value as a displayable tier; anything else falls back to 50. */
 export function resolveOverviewTier(raw: string | string[] | undefined): OverviewTier {
   const candidate = Number(Array.isArray(raw) ? raw[0] : raw);
   return OVERVIEW_TIERS.find((tier) => tier === candidate) ?? OVERVIEW_PRIMARY_TIER;
@@ -25,22 +22,13 @@ export interface OverviewTierValue {
   tier: number;
   value: number | null;
   boundary: TcoTierBoundary;
-  /**
-   * Run dates of the frontier point(s) backing this value — the two bracketing
-   * points when interpolated (from = earlier, to = later; equal on the same
-   * day), the single point twice when clamped, null when `value` is null. Comes
-   * only from this config's own frontier at this tier, never a sibling or cohort.
-   */
+  /** Bracketing frontier points when interpolated, the single point twice when
+   *  clamped; always from this config's own frontier, never a sibling's. */
   evidenceDate: { from: string; to: string } | null;
 }
 
-/**
- * One real deployable serving configuration, identified by its exact
- * deployment topology (model × hardware × framework × precision × spec_method ×
- * disagg × multinode × per-role parallelism × GPU counts × offload). Tier
- * values come from this configuration's own Pareto frontier only — never
- * blended across configurations.
- */
+/** One deployable serving configuration (exact topology identity); tier values
+ *  come from its own Pareto frontier only, never blended across configs. */
 export interface OverviewConfigResult {
   key: string;
   dbModel: string;
@@ -57,28 +45,17 @@ export interface OverviewConfigResult {
   latestDate: string;
 }
 
-/** One hardware's frontier read at a single tier, with its backing config. */
 export interface OverviewTierRead {
   tier: number;
   value: number | null;
   boundary: TcoTierBoundary | null;
-  /** Copied from the backing config's tier value; null when there is no read. */
   evidenceDate: { from: string; to: string } | null;
   config: OverviewConfigResult | null;
 }
 
-/**
- * Why a pair member shows `∞` instead of a value, from most to least
- * fundamental gap:
- *
- *  - `no_8k1k_data` — no 8K/1K single-turn row for this hardware at all;
- *  - `int4_bf16_only` — measured only outside FP4/FP8;
- *  - `standard_decode_only` — FP4/FP8 measured, but standard decode only;
- *  - `cannot_reach_at_tier` — speculative frontier tops out below the
- *    displayed tier;
- *  - `no_exact_at_tier` — no exact read at the displayed tier for any other
- *    reason (an under-swept frontier that starts above the tier).
- */
+/** Why a member shows `∞`. The subtle pair: `cannot_reach_at_tier` = every
+ *  speculative stack tops out below the tier; `no_exact_at_tier` = merely
+ *  under-swept. */
 export type OverviewMissingReason =
   | 'standard_decode_only'
   | 'int4_bf16_only'
@@ -98,18 +75,10 @@ export type OverviewHeadlinePairId = (typeof OVERVIEW_HEADLINE_PAIR_DEFINITIONS)
 export interface OverviewHeadlinePairMember {
   hardware: string;
   hardwareLabel: string;
-  /**
-   * The precision of this platform's own best qualified speculative read at
-   * the displayed tier, selected independently per platform (exact first, then
-   * value, tie → FP4). Null when the platform has no FP4/FP8 speculative
-   * bucket at all.
-   */
   precision: string | null;
-  /** The dbModel (point release) backing this platform's selected read. */
   dbModel: string | null;
-  /** Exact read when available; non-exact reads retain their boundary but never a numeric value. */
   read: OverviewTierRead;
-  /** Best @100 read from this platform's own selected dbModel and precision. */
+  /** @100 read from the same selected bucket as `read`. */
   highRead: OverviewTierRead;
   missingReason: OverviewMissingReason | null;
 }
@@ -117,19 +86,13 @@ export interface OverviewHeadlinePairMember {
 export interface OverviewHeadlinePairComparison {
   id: OverviewHeadlinePairId;
   label: string;
-  /** The precision shared by both displayed reads when the delta is comparable; null otherwise. */
   precision: string | null;
-  /** The dbModel shared by both displayed reads when the delta is comparable; null otherwise. */
   dbModel: string | null;
   candidate: OverviewHeadlinePairMember;
   baseline: OverviewHeadlinePairMember;
-  /**
-   * Candidate relative to B200. Non-null only when both displayed-tier reads
-   * are exact, share one precision AND one dbModel, and B200 > 0 — a delta
-   * never mixes FP4 with FP8 or two point releases.
-   */
+  /** Non-null only for exact reads sharing precision AND dbModel — never FP4
+   *  vs FP8 or cross-release. */
   directDeltaPercent: number | null;
-  /** Why two displayed exact reads still carry no delta; null when a delta exists or a side is missing. */
   deltaUnavailableReason: 'precision_mismatch' | 'version_mismatch' | null;
   highLeaderTransition: 'same_hardware' | 'changed_hardware' | null;
 }
@@ -137,21 +100,18 @@ export interface OverviewHeadlinePairComparison {
 export interface OverviewModelSummary {
   model: Model;
   modelLabel: string;
-  /** Fixed candidate-vs-B200 serving-stack comparisons, one per matrix column. */
   headlinePairs: OverviewHeadlinePairComparison[];
 }
 
 export interface OverviewPageData {
   models: OverviewModelSummary[];
   datasetThroughDate: string | null;
-  /** The service point every headline pair was read at. */
   tier: OverviewTier;
 }
 
-/** Precisions the overview may rank, in preference order. */
+/** In preference order — FP4 wins ties. */
 const OVERVIEW_PRECISIONS: readonly string[] = [Precision.FP4, Precision.FP8];
 
-/** Rows measuring the fixed overview workload, before any other filter. */
 function overviewWorkloadRows(rows: readonly BenchmarkRow[]): BenchmarkRow[] {
   return rows.filter(
     (row) =>
@@ -161,11 +121,8 @@ function overviewWorkloadRows(rows: readonly BenchmarkRow[]): BenchmarkRow[] {
   );
 }
 
-/**
- * Newest raw workload row across the page. Deliberately not derived from the
- * retained winners: a precision or engine the page never ranks still dates the
- * dataset it was measured in.
- */
+/** Deliberately from raw rows, not retained winners: an unranked precision or
+ *  engine still dates the dataset it was measured in. */
 export function overviewDatasetThroughDate(rows: readonly BenchmarkRow[]): string | null {
   return overviewWorkloadRows(rows).reduce<string | null>(
     (latest, row) => (latest === null || row.date > latest ? row.date : latest),
@@ -173,11 +130,7 @@ export function overviewDatasetThroughDate(rows: readonly BenchmarkRow[]): strin
   );
 }
 
-/**
- * Speculative configs for one precision, grouped by exact deployment identity.
- * Standard-decode rows never enter the pool, so a precision is ranked on its
- * speculative frontier alone.
- */
+/** Speculative-only configs for one precision, grouped by exact deployment identity. */
 function buildPrecisionConfigs(
   model: Model,
   workloadRows: readonly BenchmarkRow[],
@@ -211,16 +164,12 @@ function readConfigAtTier(config: OverviewConfigResult, tier: number): OverviewT
   };
 }
 
-/** A tier read known to be backed by a configuration. */
 interface ConfigTierRead extends OverviewTierRead {
   config: OverviewConfigResult;
 }
 
-/**
- * A tier read counts only when it lands inside the configuration's own measured
- * frontier. A clamped or unreachable read is a coverage gap, so it can never
- * lead a tier or anchor a percentage gap.
- */
+/** In-range reads only: a clamped/unreachable read is a coverage gap and never
+ *  leads a tier or anchors a delta. */
 const isExactTierRead = <T extends OverviewTierRead>(read: T): read is T & { value: number } =>
   read.value !== null && read.boundary === 'interpolated';
 
@@ -232,11 +181,8 @@ function compareTierReads(a: ConfigTierRead, b: ConfigTierRead): number {
   );
 }
 
-/**
- * One read per hardware at `tier`: its best exact in-range read, or — only when
- * the hardware has no exact read at all — its best out-of-range read, kept so
- * the hardware still surfaces as a coverage gap rather than disappearing.
- */
+/** Best exact read per hardware; a hardware with no exact read keeps its best
+ *  out-of-range read so it surfaces as a gap instead of disappearing. */
 function readsByHardwareAtTier(
   configs: readonly OverviewConfigResult[],
   tier: number,
@@ -360,16 +306,14 @@ function buildHeadlinePairs(
   workloadRows: readonly BenchmarkRow[],
   tier: OverviewTier,
 ): OverviewHeadlinePairComparison[] {
-  // FP4 and FP8 configs are built independently and never blend within a read.
   const configsByPrecision = new Map(
     OVERVIEW_PRECISIONS.map(
       (precision) => [precision, buildPrecisionConfigs(model, workloadRows, precision)] as const,
     ),
   );
-  // Each platform selects its own best qualified bucket (dbModel × precision):
-  // exact at the displayed tier first, then value, tie → FP4, newest evidence,
-  // lexical dbModel. One bucket per member keeps point releases from blending
-  // within a read.
+  // Per-platform best bucket (dbModel × precision): exact first, then value,
+  // tie → FP4, newest evidence, lexical dbModel — one bucket per member so
+  // point releases never blend within a read.
   const buildMember = (memberHardware: string): OverviewHeadlinePairMember => {
     const candidates = buildHeadlinePairBuckets(
       configsByPrecision,
@@ -440,8 +384,6 @@ function buildHeadlinePairs(
             : 'version_mismatch'
           : 'precision_mismatch'
         : null,
-      // On the 100 view the displayed reads ARE the high reads, so a
-      // transition line would restate the cell values.
       highLeaderTransition:
         comparable && tier !== OVERVIEW_HIGH_TIER
           ? headlineLeaderTransition(
@@ -509,13 +451,8 @@ export function buildOverviewModelSummary(
   };
 }
 
-/**
- * Assemble the whole page from rows already grouped by DISPLAY model. Iterating
- * DEFAULT_MODELS fixes the output order and renders every active model — a model
- * with no rows still gets its four pairs, every member carrying a missing
- * reason. The live server path and the e2e fixture path feed this same
- * function; only the row source differs.
- */
+/** DEFAULT_MODELS fixes the row order; a rowless model still renders four
+ *  pairs with missing reasons. Live and fixture paths both feed this. */
 export function assembleOverviewPageData(
   rowsByModel: Record<string, BenchmarkRow[]>,
   tier: OverviewTier = OVERVIEW_PRIMARY_TIER,
diff --git a/packages/app/src/lib/overview-links.ts b/packages/app/src/lib/overview-links.ts
index 132def11..ea8e34d1 100644
--- a/packages/app/src/lib/overview-links.ts
+++ b/packages/app/src/lib/overview-links.ts
@@ -1,10 +1,3 @@
-/**
- * @file overview-links.ts
- * @description Evidence links for the overview scorecard: a pre-filtered
- * /inference view and the CI run a configuration came from. Pure and
- * React-free so the server-rendered overview page can build them.
- */
-
 import { runIdFromRunUrl } from './known-issues';
 import {
   OVERVIEW_PRIMARY_TIER,

From b507220ec84cac2acbf1923957903a3c7a1baa16 Mon Sep 17 00:00:00 2001
From: Wenyao Gao <wgao11@u.rochester.edu>
Date: Thu, 23 Jul 2026 11:22:20 -0700
Subject: [PATCH 12/13] fix(overview): leader line follows the 100 view's own
 re-selection

---
 packages/app/src/lib/overview-data.test.ts | 15 ++++++++
 packages/app/src/lib/overview-data.ts      | 44 +++++++++++++---------
 2 files changed, 42 insertions(+), 17 deletions(-)

diff --git a/packages/app/src/lib/overview-data.test.ts b/packages/app/src/lib/overview-data.test.ts
index da4f03d7..07f2832f 100644
--- a/packages/app/src/lib/overview-data.test.ts
+++ b/packages/app/src/lib/overview-data.test.ts
@@ -294,6 +294,21 @@ describe('overview headline pairs', () => {
     expect(pair?.highLeaderTransition).toBe('changed_hardware');
   });
 
+  it('suppresses the leader line when another precision wins the 100 view', () => {
+    // Both sides are FP8-comparable @50, but the candidate's FP4 stack wins
+    // @100 — the 100 view re-selects FP4 vs FP8 and refuses to compare.
+    const summary = buildOverviewModelSummary(Model.Qwen3_5, [
+      ...frontier([1100, 900, 700, 500], { hardware: 'mi355x', precision: Precision.FP8 }),
+      ...frontier([1000, 800, 750, 700], { hardware: 'mi355x', precision: Precision.FP4 }),
+      ...frontier([1050, 850, 720, 600], { hardware: 'b200', precision: Precision.FP8 }),
+    ]);
+
+    const pair = headlinePairOf(summary, 'mi355x-vs-b200');
+    expect(pair?.directDeltaPercent).toBeCloseTo((900 / 850 - 1) * 100);
+    expect(pair?.candidate.highRead.config?.precision).toBe(Precision.FP4);
+    expect(pair?.highLeaderTransition).toBeNull();
+  });
+
   it('claims cannot-reach only when every speculative bucket is unreachable', () => {
     const unreachable: [number, number][] = [
       [20, 500],
diff --git a/packages/app/src/lib/overview-data.ts b/packages/app/src/lib/overview-data.ts
index f74d3fb0..0c2c47ca 100644
--- a/packages/app/src/lib/overview-data.ts
+++ b/packages/app/src/lib/overview-data.ts
@@ -78,7 +78,7 @@ export interface OverviewHeadlinePairMember {
   precision: string | null;
   dbModel: string | null;
   read: OverviewTierRead;
-  /** @100 read from the same selected bucket as `read`. */
+  /** The 100 view's own re-selected read, so the leader line can never disagree with ?tier=100. */
   highRead: OverviewTierRead;
   missingReason: OverviewMissingReason | null;
 }
@@ -286,11 +286,17 @@ function headlineLeaderTransition(
   candidateHigh: OverviewTierRead,
   baselineHigh: OverviewTierRead,
 ): OverviewHeadlinePairComparison['highLeaderTransition'] {
+  // A cross-precision or cross-release @100 pair carries no leader claim,
+  // mirroring the delta rule.
   if (
     !isExactTierRead(candidatePrimary) ||
     !isExactTierRead(baselinePrimary) ||
     !isExactTierRead(candidateHigh) ||
     !isExactTierRead(baselineHigh) ||
+    candidateHigh.config === null ||
+    baselineHigh.config === null ||
+    candidateHigh.config.precision !== baselineHigh.config.precision ||
+    candidateHigh.config.dbModel !== baselineHigh.config.dbModel ||
     candidatePrimary.value === baselinePrimary.value ||
     candidateHigh.value === baselineHigh.value
   ) {
@@ -314,14 +320,14 @@ function buildHeadlinePairs(
   // Per-platform best bucket (dbModel × precision): exact first, then value,
   // tie → FP4, newest evidence, lexical dbModel — one bucket per member so
   // point releases never blend within a read.
-  const buildMember = (memberHardware: string): OverviewHeadlinePairMember => {
+  const selectRead = (memberHardware: string, atTier: OverviewTier) => {
     const candidates = buildHeadlinePairBuckets(
       configsByPrecision,
       new Set([memberHardware]),
-      tier,
+      atTier,
     ).map((memberBucket) => ({
       bucket: memberBucket,
-      read: nonComparableAsMissing(memberBucket.tierReads.get(memberHardware), tier),
+      read: nonComparableAsMissing(memberBucket.tierReads.get(memberHardware), atTier),
     }));
     const best = candidates.toSorted(
       (a, b) =>
@@ -332,25 +338,29 @@ function buildHeadlinePairs(
         b.bucket.newestEvidence.localeCompare(a.bucket.newestEvidence) ||
         a.bucket.dbModel.localeCompare(b.bucket.dbModel),
     )[0];
-    const primary = best?.read ?? nullTierRead(tier);
-    const high = nonComparableAsMissing(
-      best === undefined
-        ? undefined
-        : readsByHardwareAtTier(best.bucket.configs, OVERVIEW_HIGH_TIER).get(memberHardware),
-      OVERVIEW_HIGH_TIER,
-    );
+    return {
+      reads: candidates.map(({ read }) => read),
+      bucket: best?.bucket ?? null,
+      read: best?.read ?? nullTierRead(atTier),
+    };
+  };
+
+  const buildMember = (memberHardware: string): OverviewHeadlinePairMember => {
+    const primary = selectRead(memberHardware, tier);
+    const high =
+      tier === OVERVIEW_HIGH_TIER ? primary : selectRead(memberHardware, OVERVIEW_HIGH_TIER);
     return {
       hardware: memberHardware,
       hardwareLabel: getHardwareConfig(memberHardware, model).label,
-      precision: best?.bucket.precision ?? null,
-      dbModel: best?.bucket.dbModel ?? null,
-      read: primary,
-      highRead: high,
+      precision: primary.bucket?.precision ?? null,
+      dbModel: primary.bucket?.dbModel ?? null,
+      read: primary.read,
+      highRead: high.read,
       missingReason: missingReasonForHeadlineMember(
         workloadRows,
         memberHardware,
-        primary,
-        candidates.map(({ read }) => read),
+        primary.read,
+        primary.reads,
       ),
     };
   };

From 5aade7294ebe86422c47530fcd1ac0c449a6323c Mon Sep 17 00:00:00 2001
From: Wenyao Gao <wgao11@u.rochester.edu>
Date: Thu, 23 Jul 2026 11:22:33 -0700
Subject: [PATCH 13/13] fix(header): carry query state into the language toggle
 href

---
 packages/app/cypress/e2e/overview.cy.ts       |  4 +++-
 packages/app/src/components/header/header.tsx | 12 ++++++++++--
 2 files changed, 13 insertions(+), 3 deletions(-)

diff --git a/packages/app/cypress/e2e/overview.cy.ts b/packages/app/cypress/e2e/overview.cy.ts
index 2219a98e..955f1249 100644
--- a/packages/app/cypress/e2e/overview.cy.ts
+++ b/packages/app/cypress/e2e/overview.cy.ts
@@ -260,7 +260,9 @@ describe('Overview page', () => {
 
     cy.visit('/overview?tier=100');
     cy.contains('Output tok/s/GPU @100 tok/s/user').should('exist');
-    cy.get('[data-testid="language-toggle"]').click();
+    cy.get('[data-testid="language-toggle"]')
+      .should('have.attr', 'href', '/zh/overview?tier=100')
+      .click();
     cy.location('pathname').should('eq', '/zh/overview');
     cy.location('search').should('eq', '?tier=100');
     cy.contains('每 GPU 输出 tok/s @100 tok/s/用户').should('exist');
diff --git a/packages/app/src/components/header/header.tsx b/packages/app/src/components/header/header.tsx
index bc438d34..021d5795 100644
--- a/packages/app/src/components/header/header.tsx
+++ b/packages/app/src/components/header/header.tsx
@@ -85,15 +85,23 @@ function LanguageToggle({
 }) {
   const isZh = isZhPathname(pathname);
   const target = switchLocalePath(pathname);
+  // The href carries the query too, so modified clicks and copied links keep
+  // shareable state (e.g. the overview's ?tier=) — same sync as TabNav's.
+  const [search, setSearch] = useState('');
+  useEffect(() => {
+    const sync = () => setSearch(window.location.search);
+    sync();
+    window.addEventListener('popstate', sync);
+    return () => window.removeEventListener('popstate', sync);
+  }, [pathname]);
   return (
     <Link
-      href={target}
+      href={target + search}
       data-testid="language-toggle"
       hrefLang={isZh ? 'en' : 'zh-CN'}
       className="inline-flex items-center min-h-11 px-2 rounded-md text-sm font-medium text-muted-foreground hover:text-foreground hover:bg-muted transition-colors whitespace-nowrap"
       onClick={(event) => {
         track('header_language_toggled', { to: isZh ? 'en' : 'zh' });
-        // The sibling keeps shareable query state (e.g. the overview's ?tier=).
         navigateInApp(event, router, target + window.location.search);
       }}
     >