calculator: add liability disclaimer + rename Production->Deployment (Alex review)

Per review feedback on PR #34:
- Add a free-utility / no-warranty / no-liability / AS-IS disclaimer to the
  README, both web pages (visible banner), and the calc.py header.
- Rename PRODUCTION-GUIDE.md -> DEPLOYMENT-GUIDE.md; remove the word
  'Production' from all content (-> deployment / live environment).
- Clean dangling driver/sweep.py refs in shipped files -> 'a live GPU sweep'
  (the sweep harness lives in the companion repo, noted in the scope banner).

Author: Anton Alexander

Author: dmvevents

Container-Root/eks/deployment/inference/agentic-ai/disagg-calculator/README.md

@@ -2,6 +2,14 @@
 
 Author: Anton Alexander
 
+> **Disclaimer.** This calculator is provided **free of charge, as a utility** to help quickly
+> **estimate** and plan model-serving topologies. It produces **first-order estimates and guidance only**.
+> AWS and the authors make **no representation, warranty, or guarantee** — express or implied — about model
+> performance, throughput, latency, cost, or fitness for any purpose, and **accept no liability** for any
+> outcome arising from its use. Actual measured results will differ from the calculated numbers. Always
+> validate on your own hardware with your own workload before making deployment or purchasing decisions.
+> The calculator is provided **"AS IS"** without support obligation.
+
 An architecture-first calculator that answers a single question for any model on any GPU:
 **should I serve it aggregated (prefill + decode on one worker) or disaggregated (prefill and decode
 split across workers, KV cache moved over the network)?** It derives the answer from the model's real
@@ -28,7 +36,7 @@ disagg-calculator/
     index.html       # interactive calculator (pick a model + GPU + workload → verdict), runs in-browser
     calculator.html  # the same engine as a single-purpose page (shareable link)
   docs/
-    PRODUCTION-GUIDE.md  # when to disaggregate vs. scale aggregated replicas — rules of thumb + worked cases
+    DEPLOYMENT-GUIDE.md  # when to disaggregate vs. scale aggregated replicas — rules of thumb + worked cases
     METHODOLOGY.md       # the math, the physics assumptions, and how the calibration loop refines them
 ```
 
@@ -77,7 +85,7 @@ accurate **without changing the physics**. See [docs/METHODOLOGY.md](docs/METHOD
 
 ## When to use which verdict
 
-See [docs/PRODUCTION-GUIDE.md](docs/PRODUCTION-GUIDE.md) for the rules of thumb. In short:
+See [docs/DEPLOYMENT-GUIDE.md](docs/DEPLOYMENT-GUIDE.md) for the rules of thumb. In short:
 
 - **Aggregated replicas** when: short-context chat, prefill cheap relative to decode, cost/throughput-bound,
   loose-to-moderate ITL SLO. At the 1P:1D floor, disaggregation buys little but costs ≥2× the GPUs.

Container-Root/eks/deployment/inference/agentic-ai/disagg-calculator/analyze/calc.py

@@ -3,6 +3,10 @@
 Author: Anton Alexander
 Source: github.com/dmvevents/dynamo-disagg-optimizer · live: dmvevents.github.io/dynamo-disagg-optimizer/calculator.html
 
+DISCLAIMER: Provided free, as a utility, for estimation and planning ONLY. AWS and the authors make no
+warranty or guarantee of model performance and accept no liability for results obtained from its use.
+Actual measured numbers will differ; validate on your own hardware before deployment decisions. AS IS.
+
 calc.py — analytical "should I disaggregate?" model for ANY (instance, model, workload).
 
 Given an AWS instance type + model characteristics + workload SLO, estimate — WITHOUT running GPUs —
@@ -12,7 +16,7 @@
 
 Grounded in first-order rooflines and CALIBRATED to this org's measured Nemotron-3 Ultra 550B / p5en
 numbers (see CAL/calibration below). It is an ESTIMATOR (order-of-magnitude + direction), not a simulator —
-it tells you where to spend GPU hours, then the live sweep (driver/sweep.py) confirms.
+it tells you where to spend GPU hours, then the live sweep (a live GPU sweep) confirms.
 
 Pure-Python stdlib only, so it runs headless and the HTML pages reimplement the SAME formulas in JS
 (kept honest by tests/parity.py, which asserts calc.py == calculator.html == index.html on a grid).
@@ -496,12 +500,12 @@ def _caveats(model: dict, itl_cal: dict | None = None) -> list[str]:
     if itl_cal is not None:
         if itl_cal.get("calibration_n", 0) >= 1:
             out.append("decode ITL is calibrated to a MEASURED anchor (n=%d, %s); raw roofline shown alongside. "
-                       "Single-point calibration — re-confirm off the bf16/H200/TP8 anchor with driver/sweep.py."
+                       "Single-point calibration — re-confirm off the bf16/H200/TP8 anchor with a live GPU sweep."
                        % (itl_cal["calibration_n"], itl_cal.get("calibration_confidence", "")))
         else:
             out.append("decode ITL is the RAW HBM roofline (no measured anchor for this family) — likely OPTIMISTIC "
                        "for eager-mode / low-concurrency; treat as a lower bound on latency.")
-    out.append("Estimator only: directionally accurate, calibrated to 550B/p5en (n=1). Confirm the verdict with driver/sweep.py on real GPUs.")
+    out.append("Estimator only: directionally accurate, calibrated to 550B/p5en (n=1). Confirm the verdict with a live GPU sweep on real GPUs.")
     return out
 
 

…sagg-calculator/docs/PRODUCTION-GUIDE.md …sagg-calculator/docs/DEPLOYMENT-GUIDE.mdContainer-Root/eks/deployment/inference/agentic-ai/disagg-calculator/docs/PRODUCTION-GUIDE.md renamed to Container-Root/eks/deployment/inference/agentic-ai/disagg-calculator/docs/DEPLOYMENT-GUIDE.md Container-Root/eks/deployment/inference/agentic-ai/disagg-calculator/docs/PRODUCTION-GUIDE.md renamed to Container-Root/eks/deployment/inference/agentic-ai/disagg-calculator/docs/DEPLOYMENT-GUIDE.md

@@ -1,4 +1,4 @@
-# Production guide: aggregation vs disaggregation, end-to-end
+# Deployment guide: aggregation vs disaggregation, end-to-end
 
 *Author: Anton Alexander*  
 *Source: github.com/dmvevents/dynamo-disagg-optimizer · live: dmvevents.github.io/dynamo-disagg-optimizer/calculator.html*
@@ -7,7 +7,7 @@
 A decision-then-deploy runbook. You start with a model + an SLA + a budget, and finish with a deployed,
 benchmarked Dynamo serving config — having spent GPU hours **only where the cheap analytical layers said to**.
 
-The question this guide answers in production terms: **for MY model, instance, SLA, and budget — do I
+The question this guide answers in deployment terms: **for MY model, instance, SLA, and budget — do I
 aggregate or disaggregate prefill/decode, and what does each cost per million tokens?**
 
 > Latency and cost can disagree. Disaggregation often wins on **TTFT / SLA-attainment** while
@@ -119,10 +119,10 @@ python3 analyze/cost.py --instance p5en.48xlarge --config m.json \
    --price-per-hr <your_rate_if_capacity_block>      # default = live AWS on-demand
 ```
 Output: $/Mtok for agg vs disagg, cheaper-at-SLA, savings %, and the latency-vs-cost note. **This is the
-production decision artifact** — pair it with the SLA-attainment from step 6.
+deployment decision artifact** — pair it with the SLA-attainment from step 6.
 
 ## 8. DEPLOY the winner
-`kubectl apply -f <chosen YAML>` (or the aiconfigurator DGD). Re-run step 6 in prod periodically; pod-freshness
+`kubectl apply -f <chosen YAML>` (or the aiconfigurator DGD). Re-run step 6 in your live environment periodically; pod-freshness
 matters (a stale pod can show 4× worse latency — restart before trusting a baseline).
 
 ---

Container-Root/eks/deployment/inference/agentic-ai/disagg-calculator/web/calculator.html

@@ -116,6 +116,13 @@ <h1>Aggregation vs Disaggregation Calculator</h1>
    transfer-toll vs interference → <b>verdict</b> → $/Mtok → optimal operating point.
    <b class="honest">Disaggregation is not free — it helps for some workloads and regresses for others. This tool tells you which: when it helps, when it&rsquo;s a no-op, and when it costs you (at the 1P:1D floor it&rsquo;s a latency play, not a cost play).</b>
    Physics mirrors <code>analyze/calc.py</code> + <code>cost.py</code> + <code>optimizer.py</code>, calibrated to measured 550B/p5en.</div>
+  <div style="margin:10px 0;padding:8px 12px;border:1px solid #c89b3c;background:#2a2410;color:#e8d9a0;border-radius:6px;font-size:12px;line-height:1.45">
+    <b>Disclaimer.</b> Provided free of charge, as a utility, for quick <b>estimation and planning only</b>. AWS and the
+    authors make <b>no representation, warranty, or guarantee</b> of model performance, throughput, latency, or cost, and
+    <b>accept no liability</b> for any outcome arising from its use. Actual measured results will differ from the
+    calculated numbers — always validate on your own hardware with your own workload before deployment or purchasing
+    decisions. Provided <b>&ldquo;AS IS&rdquo;</b>.
+  </div>
 </header>
 <div class="wrap">
   <!-- ============ INPUTS (the architecture) ============ -->
@@ -205,7 +212,7 @@ <h3>2 · Compute + workload + targets</h3>
         <b>How to read this:</b> a <b>directional L0 screen</b>, calibrated to a single measured anchor
         (<b>n=1</b>): Nemotron-3 Ultra 550B, <b>bf16</b>, <b>H200 / p5en</b>, <b>TP8</b>. It models per-request
         single-stream latency — <b>not</b> a validated multi-tenant simulator. Confirm any verdict with
-        <code>driver/sweep.py</code> on real GPUs before acting.
+        <code>a live GPU sweep</code> on real GPUs before acting.
         <div class="provlegend">Number provenance:
           <span class="badge meas tiny">MEASURED</span> real GPU benchmark
           <span class="badge der tiny">DERIVED</span> computed from your inputs

Container-Root/eks/deployment/inference/agentic-ai/disagg-calculator/web/index.html

@@ -139,7 +139,14 @@
 <body>
 <header>
   <h1>Aggregation vs Disaggregation Optimizer <span class="brand">— should you split prefill/decode?</span></h1>
-  <div class="sub">Enter an AWS instance + model + workload → get the <b>topology verdict</b>: aggregate, hybrid, or disaggregate. <b class="honest">Disaggregation is not free — it helps for some workloads and regresses for others. This tool tells you which: when it helps, when it&rsquo;s a no-op, and when it costs you (at the 1P:1D floor it&rsquo;s a latency play, not a cost play).</b> Physics mirrors <code>analyze/calc.py</code>; calibrated to measured Nemotron-3 Ultra 550B / p5en. Estimator — confirm with <code>driver/sweep.py</code>.</div>
+  <div class="sub">Enter an AWS instance + model + workload → get the <b>topology verdict</b>: aggregate, hybrid, or disaggregate. <b class="honest">Disaggregation is not free — it helps for some workloads and regresses for others. This tool tells you which: when it helps, when it&rsquo;s a no-op, and when it costs you (at the 1P:1D floor it&rsquo;s a latency play, not a cost play).</b> Physics mirrors <code>analyze/calc.py</code>; calibrated to measured Nemotron-3 Ultra 550B / p5en. Estimator — confirm with a live sweep.</div>
+  <div style="margin:10px 0;padding:8px 12px;border:1px solid #c89b3c;background:#2a2410;color:#e8d9a0;border-radius:6px;font-size:12px;line-height:1.45">
+    <b>Disclaimer.</b> Provided free of charge, as a utility, for quick <b>estimation and planning only</b>. AWS and the
+    authors make <b>no representation, warranty, or guarantee</b> of model performance, throughput, latency, or cost, and
+    <b>accept no liability</b> for any outcome arising from its use. Actual measured results will differ from the
+    calculated numbers — always validate on your own hardware with your own workload before deployment or purchasing
+    decisions. Provided <b>&ldquo;AS IS&rdquo;</b>.
+  </div>
 </header>
 
 <!-- ============ EXPLAINER ============ -->
@@ -391,7 +398,7 @@ <h4>2. Prefill/Decode pool scaling — the at-scale agg-vs-disagg crossover</h4>
         <b>How to read this:</b> a <b>directional L0 screen</b>, calibrated to a single measured anchor
         (<b>n=1</b>): Nemotron-3 Ultra 550B, <b>bf16</b>, <b>H200 / p5en</b>, <b>TP8</b>. It models per-request
         single-stream latency — <b>not</b> a validated multi-tenant simulator. Confirm any verdict with
-        <code>driver/sweep.py</code> on real GPUs before acting.
+        <code>a live GPU sweep</code> on real GPUs before acting.
       </div>
       <div class="trustchip"><span class="dot" aria-hidden="true"></span><span>Honest by design — across our 240-cell catalog the verdict splits roughly evenly: <b>~52% disaggregate, ~48% aggregate-or-hybrid</b>. And on the cost axis at the 1P:1D floor, disagg almost always regresses (it&rsquo;s a latency play). A no-op or a regression is a real result, not a failure.</span></div>
       <details class="whatis">
@@ -710,7 +717,7 @@ <h4>✗ It is NOT</h4>
  document.getElementById('why').textContent=r.why;
  document.getElementById('vmodel').textContent=(preset.value||"custom")+" ("+m.family+")";
  document.getElementById('vinst').textContent=document.getElementById('inst').value+" ("+g.gpu+")";
- document.getElementById('cav').textContent = m.family==="hybrid-mamba"?"⚠ hybrid-Mamba: PP1 only on EFA; KV pull descriptor-bound. Confirm with driver/sweep.py.":"Estimator — confirm verdict with driver/sweep.py on real GPUs.";
+ document.getElementById('cav').textContent = m.family==="hybrid-mamba"?"⚠ hybrid-Mamba: PP1 only on EFA; KV pull descriptor-bound. Confirm with a live GPU sweep.":"Estimator — confirm verdict with a live GPU sweep on real GPUs.";
  kKV.textContent=r.kv_MB.toFixed(1)+" MB"; kXfer.textContent=r.xfer.toFixed(1)+" ms"; kPre.textContent=r.pre.toFixed(0)+" ms";
  kITL.textContent=r.itl.toFixed(1)+" ms"; kLine.textContent=r.line.toFixed(0)+" GB/s"; kImb.textContent=r.imb.toFixed(1)+"×";
  bar('cTime',['KV xfer','Prefill','Decode/tok','Decode phase'],[r.xfer,r.pre,r.itl,r.dphase],['#58a6ff','#d29922','#3fb950','#8957e5'],'Latency components (ms)');