release: v3.15.0 — E1 v3 verification campaign + arXiv-ready papers + BEAM-10M harness

Single coherent docs + metadata commit cutting v3.15.0. 64 commits since
v3.14.12 grouped thematically in CHANGELOG. README updated minimally:

- Version badge 3.14.2 -> 3.15.0; citation badge 41 -> 45 papers.
- Lead intro updated: 20 -> 26 biological mechanisms (matches 26-enum
  ablation campaign in tasks/e1-v3-results.md).
- New v3.15.0 lead-line summarising the verification campaign +
  production fixes + arXiv-ready papers + BEAM-10M harness.
- LoCoMo headline R@10 94.2% -> 94.3% and MRR 0.8278 -> 0.8279 to match
  the post-plasticity-fix BASELINE_NO_CONSOLIDATION JSON
  (benchmarks/results/ablation/locomo_v3_post_plasticity_fix/).
- Science section: 41 -> 45 papers; new direct links to both PDF
  artefacts (arxiv-thermodynamic 30 pages, arxiv-context-assembly 37 pages).
- Verification section: now references both papers + 26 enum mechanisms.
- Citation block: adds @Unpublished entries for both PDFs with
  "arXiv ID forthcoming, endorsement in progress".

CHANGELOG.md: full v3.15.0 entry covering verification campaign
(LongMemEval-S 17 rows n=500, LoCoMo 14 rows × 2 sweeps n=1986, blend
calibration, per-category specialization), production fixes (consolidation
cadence ingest_at, plasticity result-shape contract), wiring
(HOPFIELD/HDC/SPREADING_ACTIVATION/DENDRITIC_CLUSTERS/EMOTIONAL_RETRIEVAL/
MOOD_CONGRUENT_RERANK/RECONSOLIDATION + 23-mechanism CORTEX_ABLATE_<MECH>
hooks at production hot-path), benchmark infrastructure (BEAM-10M LLM
head-to-head harness, LoCoMo --ablate flags, blend calibration),
papers (recompile with all 45 citations resolved, §6.3 three-pass
integration), issue fixes (#15 Nitjsefnie session-layout discovery),
and repo housekeeping.

Headline numbers verified against on-disk JSONs:
- LongMemEval-S: MRR=0.9124, R@10=98.4% (n=500)
- LoCoMo (post-plasticity-fix): MRR=0.8279, R@10=94.3% (n=1986)
- BEAM Overall: 0.591 (unchanged since 5812342)

No production code touched in this commit.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Author: cdeust

.claude-plugin/marketplace.json

@@ -6,14 +6,14 @@
   },
   "metadata": {
     "description": "Persistent memory and cognitive profiling plugins for Claude Code",
-    "version": "3.14.5"
+    "version": "3.15.0"
   },
   "plugins": [
     {
       "name": "cortex",
       "source": "./",
       "description": "Persistent memory and cognitive profiling for Claude Code — thermodynamic memory with heat/decay, intent-aware retrieval, biological plasticity, codebase intelligence, and cognitive profiling. 47 MCP tools with enriched schemas. PostgreSQL + pgvector in CLI mode; automatic SQLite fallback in Cowork/sandboxed mode. Curated wiki (ADRs, specs, lessons) with audit-artefact filtering. Consolidate is set-based SQL batched — decay/plasticity/pruning run 100-500× faster on large stores. Workflow graph with caller-qualified CALLS chains rendering full method-to-method dependencies (native tree-sitter, no AP required). Side panel humanized for non-technical users. Ingests codebase analysis (ai-automatised-pipeline) and PRDs (prd-spec-generator) into wiki + memory + knowledge graph. Docker image available.",
-      "version": "3.14.5",
+      "version": "3.15.0",
       "author": {
         "name": "Clement Deust",
         "email": "admin@ai-architect.tools"

CHANGELOG.md

@@ -6,6 +6,153 @@ adheres to [Semantic Versioning](https://semver.org/).
 
 ## [Unreleased]
 
+## [3.15.0] — E1 v3 verification campaign + arXiv-ready papers + BEAM-10M harness
+
+A single coherent release covering 64 commits since v3.14.12. The headline is
+verification: every benchmark number on the README is now backed by a
+per-mechanism ablation row with code SHAs, dirty flags, manifests, and
+per-row JSON outputs preserved alongside the writeups. Two production fixes
+were surfaced by the campaign and ship inside the same release. Both
+companion papers (thermodynamic memory + structured context assembly) are
+arXiv-ready.
+
+### Verification campaign (paper-claim-bearing)
+
+- **E1 v3 LongMemEval-S — 17-row per-mechanism ablation, n=500.** Headline
+  `MRR = 0.9124`, `R@10 = 98.4%` (vs. published baselines `MRR = 0.882`,
+  `R@10 = 97.8%`: **+3.0% MRR, +0.6% R@10**). Driver:
+  `benchmarks/lib/run_e1_v3_lme.py`. Per-row JSONs:
+  `benchmarks/results/ablation/longmemeval-s_v3/`. Writeup:
+  `tasks/e1-v3-results.md`.
+- **E1 v3 LoCoMo — 14-row two-baseline ablation, n=1986.** Headline
+  `MRR = 0.8279`, `R@10 = 94.3%` (`BASELINE_NO_CONSOLIDATION`,
+  longitudinal-read-path anchor) — vs. CLAUDE.md baseline
+  (`MRR = 0.794`, `R@10 = 0.926`): **+4.3% MRR, +1.7% R@10**. Re-run on
+  plasticity-fixed bytes (commit `2f45bcb`, descendant of `5f737fe`).
+  Cadence-fix anchor agreement re-validated identically
+  (`ΔvsNO = +0.0014`); two consolidation-only rows
+  (`HOMEOSTATIC_PLASTICITY`, `SCHEMA_ENGINE`) recover positive
+  contributions previously masked by the contract bug.
+  `benchmarks/results/ablation/locomo_v3_post_plasticity_fix/`.
+  Writeup: `tasks/e1-v3-locomo-results-post-fix.md`. The pre-fix sweep is
+  preserved at `tasks/e1-v3-locomo-results.md`.
+- **Phase A + B blend-weight calibration.** Central composite design + 5×5
+  grid search; all six post-WRRF rerank constants confirmed near-optimum at
+  the engineering defaults shipped today. `tasks/e1-v3-blend-calibration.md`.
+- **Per-category delta analysis (LME-S).** Mechanism specialization
+  surfaced: HDC specializes for multi-session reasoning, HOPFIELD for
+  knowledge updates, ADAPTIVE_DECAY against stable preferences.
+  `tasks/e1-v3-per-category.md`.
+
+Total: **45 per-mechanism evidence rows** across 26 enum mechanisms
+(17 read-path on LongMemEval-S + 9 consolidation-only routed to LoCoMo).
+
+### Fixed (production fixes surfaced during verification)
+
+- **`6c51bce` — consolidation cadence is now ingest-relative.**
+  `consolidation_engine` migrated from wall-clock `created_at` to
+  ingest-relative `ingested_at`. Recovers `MRR 0.222 → 0.8264` on
+  backdated corpora; affects every production backfill scenario where
+  memories carry old timestamps but were written today.
+- **`5f737fe` — plasticity result-shape contract preserved on ablation.**
+  `apply_hebbian_update` no-op (when `CORTEX_ABLATE_SYNAPTIC_PLASTICITY=1`)
+  now returns dicts with `action="none"` instead of raw edge tuples,
+  fixing a silent `KeyError` downstream in consolidation/plasticity. This
+  is what was masking the two consolidation-only contributions in the
+  pre-fix LoCoMo sweep.
+
+### Added (read-path mechanisms now wired end-to-end)
+
+- **`ddb5b58` / `024ea1a` / `bc0ae4f`** — `HOPFIELD`, `HDC`,
+  `SPREADING_ACTIVATION`, `DENDRITIC_CLUSTERS` wired into the `pg_recall`
+  pipeline. Batch Hopfield embeddings and real entity-set Jaccard for the
+  dendritic stage. Query-entity resolution extended to natural-language
+  tokens.
+- **`81e8d90`** — `EMOTIONAL_RETRIEVAL` + `MOOD_CONGRUENT_RERANK` are now
+  live read-path stages (not test-only).
+- **`9d6bc96`** — `RECONSOLIDATION` post-retrieval stage wired
+  (Nader 2000); retrieved memories become labile and may be updated
+  against the retrieval context.
+- **`c5ade6b`** — VADER → `user_mood` EMA hook in `remember()`; closes
+  the `MOOD_CONGRUENT` signal gap end-to-end.
+- **`b4b23e7`** — `PgMemoryStore.get_user_mood` / `set_user_mood` +
+  `user_mood` DDL; the column the read-path stage was reading didn't
+  exist before this.
+- **`099ba1e` / `54f8501`** — 23 mechanisms now have
+  `CORTEX_ABLATE_<MECH>=1` env-var hooks reading at the production
+  hot-path (not just at test wiring), so ablation studies exercise the
+  same code path as production.
+
+### Added (benchmark + verification infrastructure)
+
+- **`3201cc3` / `0a53996`** — BEAM-10M LLM head-to-head harness scaffold
+  + live mode wiring at `benchmarks/llm_head_to_head/`; smoke pending
+  API keys.
+- **`0e1f90d`** — LongMemEval-S `--with-consolidation` flag.
+- **`b68c5ac` / `ef178da`** — LoCoMo `--ablate` + `--with-consolidation`
+  + `--results-out` flags + 14-row driver `run_e1_v3_locomo.py`.
+- **`f09485d`** — Blend-weight calibration infrastructure with
+  pre-registration; harness dirty-check matched to pre-reg
+  (`39ab694` ignores submodule internal state).
+- **`5a5d8d3` / `3eab1ed`** — E2 N-scan rebuilt as real-benchmark
+  subsample + Zipf synthetic; ablation env vars wired into the
+  production code path.
+- **DB snapshot + restore + HNSW determinism infrastructure** (E2 / E3 /
+  E4 / E5 internal harnesses).
+
+### Added (papers + endorsement materials)
+
+- **`6b80760` / `3ace1fb` / `3eaeaf6`** — `docs/arxiv-thermodynamic/main.pdf`
+  compiled, 30 pages. Ported to LaTeX matching `arxiv/main.tex` style.
+- **`9e6ddf6`** — Recompile with bibtex pass; **all 45 citations now
+  resolve** (vs. the previous 4 unresolved `??` markers).
+- **`bce4840` / `db4fe0a` / `6f75221`** — §6.3 three-pass integration:
+  LME-S evidence + LoCoMo subsection + post-fix re-run + cadence-fix
+  narrative + plasticity-fix narrative.
+- **`fa9c101` / `fb6f67f`** — §6.4 Operating Regime added; full E2b Zipf
+  curve integrated; falsifications reframed as predicted boundaries with
+  the `N=100k` datapoint landed.
+- **`a787fe6`** — Refresh `linkedin-endorser-post.md`; new
+  `arxiv-endorsement-email.md` template with pre-submission checklist.
+- **`974c364` / `2152946`** — Prose polish; `BEAM Overall 0.543 → 0.591`
+  number fix in CLAUDE.md and the markdown source.
+- **`ffcad91`** — Repo reorg: `arxiv/` → `arxiv-context-assembly/` +
+  paper-md moved into `docs/papers/`.
+- **`docs/arxiv-context-assembly/main.pdf`** — 37 pages, pre-existing
+  verbatim + argmax bugs fixed, arXiv-ready.
+
+### Fixed (issue fixes from contributors)
+
+- **`5398745`** — issue #15 (Nitjsefnie). `discover_files` walks all four
+  session layouts (subagent + teammate transcripts), recovers ~89% of
+  session content during backfill that was previously dropped.
+
+### Fixed (CI + plumbing)
+
+- **`df14e16`** — DDL comment semicolon broke `ddl.split(';')` extractor.
+- **`9f94bd3`** — `user_mood` DDL comment semicolon + test uses dominant
+  beta.
+- **`34aa452`** — Repair docstring boundary in `cls.run_cls_cycle`
+  (broken in `3eab1ed`).
+- **`51ce608` / `c4253cc` / `5271828` / `fd51f6f` / `4918638` / `79f0b20`** —
+  ruff format + drop unused imports in verification harnesses;
+  bump tool count to 47.
+- **`18b4be4`** — ruff format on `memories_page` + `memories_facets`.
+
+### Changed (visualization, repo housekeeping)
+
+- **`63bacca` / `2953bae` / `b7a8f97`** — Paged Knowledge + Board with
+  filter chips, lazy-load; default landing reverted to Knowledge; Graph
+  view restored to pre-d3-removal state with a warning banner.
+- **22 stale public repos archived; `ai-prd-mcp` deleted** — security
+  hardening (legacy build artefacts had embedded keys at one point) +
+  portfolio cleanup.
+- **`551a411` / `30d80fe`** — Profile README draft for `cdeust/cdeust`
+  (controls AI Overview narrative); profile draft points
+  `AI Architect` to website not archived repo.
+- **Cortex repo description + topics refreshed** for AI-search
+  discovery.
+
 ## [3.14.12] — fix MCP client deadlock on long upstream responses
 
 ### Fixed

README.md

@@ -7,8 +7,8 @@
   <a href="LICENSE"><img src="https://img.shields.io/badge/License-MIT-blue.svg" alt="MIT License"></a>
   <img src="https://img.shields.io/badge/python-3.10+-blue.svg" alt="Python 3.10+">
   <img src="https://img.shields.io/badge/tests-2500_passing-brightgreen.svg" alt="Tests">
-  <img src="https://img.shields.io/badge/citations-41_papers-orange.svg" alt="Citations">
-  <img src="https://img.shields.io/badge/version-3.14.2-brightgreen.svg" alt="Version 3.14.2">
+  <img src="https://img.shields.io/badge/citations-45_papers-orange.svg" alt="Citations">
+  <img src="https://img.shields.io/badge/version-3.15.0-brightgreen.svg" alt="Version 3.15.0">
   <a href="https://glama.ai/mcp/servers/cdeust/Cortex"><img src="https://glama.ai/mcp/servers/cdeust/Cortex/badges/score.svg" alt="Glama score: security A, license A"></a>
 </p>
 
@@ -30,7 +30,9 @@ Claude Code forgets you every time you close the tab. Every architecture decisio
 
 Cortex is a persistent memory engine for Claude Code built on computational neuroscience. It remembers what you worked on, how you think, what you decided and why. Not as a dumb text dump shoved into context, but as a living memory system that consolidates, forgets intelligently, and reconstructs the right context at the right time.
 
-**20 biological mechanisms. 47 MCP tools. 9 automatic hooks. Runs entirely on your machine. PostgreSQL + pgvector.**
+**26 biological mechanisms. 47 MCP tools. 9 automatic hooks. Runs entirely on your machine. PostgreSQL + pgvector.**
+
+**v3.15.0 — verification campaign + arXiv-ready papers**: 45 per-mechanism ablation rows across LongMemEval-S (17 rows, n=500) and LoCoMo (14 rows × 2 sweeps, n=1986). Headline numbers stay verified — LongMemEval R@10 = 98.4% / MRR = 0.9124, LoCoMo R@10 = 94.3% / MRR = 0.8279 — and every figure now traces to a JSON in `benchmarks/results/ablation/` with code SHAs, dirty flags, and per-row category breakdowns. The thermodynamic memory paper (`docs/arxiv-thermodynamic/main.pdf`, 30 pages, all 45 citations resolved) and the structured context assembly paper (`docs/arxiv-context-assembly/main.pdf`, 37 pages) are arXiv-ready. Two production fixes surfaced during verification: consolidation cadence is now ingest-relative instead of wall-clock (recovers MRR 0.222 → 0.8264 on backdated corpora), and the plasticity ablation no-op preserves the result-shape contract (no more silent KeyError). HOPFIELD, HDC, SPREADING_ACTIVATION, DENDRITIC_CLUSTERS, EMOTIONAL_RETRIEVAL, MOOD_CONGRUENT_RERANK, and RECONSOLIDATION are now wired end-to-end on the production read path; 23 mechanisms have CORTEX_ABLATE_<MECH> hooks reading at the hot path. BEAM-10M LLM head-to-head harness scaffolded at `benchmarks/llm_head_to_head/`. [Release notes →](https://github.com/cdeust/Cortex/releases/tag/v3.15.0)
 
 **v3.14.2 — call graph lit + queryable**: the workflow graph now renders the actual call and import edges between symbols — not just the AST shells. Every edge carries a *confidence* (0.0–1.0) and a *reason* tag (`direct-ast`, `import-scope-lookup`, `memory-entities-link`, …) so you can tell a resolved call from a same-name guess at a glance. Knowledge-graph entities ship as a first-class layer: ~10k entities extracted from memory text land between the memory ring and the file shell, heat-weighted centroid-placed near the memories that mention them. And a new `query_workflow_graph` MCP tool returns typed subgraphs on demand — filter by `node_kind`, `edge_kind`, `neighbour_of <id> + depth`, or `domain`, so downstream agents can reason over graph slices without rebuilding from scratch.
 
@@ -152,8 +154,8 @@ LoCoMo (Maharana et al., ACL 2024): 1,986 questions across 10 conversations, inc
 
 | | Cortex | What it means |
 |---|---|---|
-| Recall@10 | **94.2%** | Right memory in top 10 over 9 times out of 10 (n=1986, BASELINE_NO_CONSOLIDATION) |
-| MRR | **0.8278** | Correct answer is typically the first result |
+| Recall@10 | **94.3%** | Right memory in top 10 over 9 times out of 10 (n=1986, BASELINE_NO_CONSOLIDATION, post-plasticity-fix) |
+| MRR | **0.8279** | Correct answer is typically the first result |
 
 | Category | MRR | R@10 | Why this score |
 |---|---|---|---|
@@ -234,7 +236,7 @@ Cortex doesn't store memories the way a database stores rows. It treats them mor
 
 **Similar memories stay distinct.** Pattern separation — modeled on the dentate gyrus, which keeps "Tuesday's standup" separate from "Wednesday's standup" even though they're almost identical. Without this, retrieval returns the same generic match for every similar query. *(Leutgeb et al. 2007; Yassa & Stark 2011)*
 
-**41 papers total.** Every algorithm, constant, and threshold traces to a published source. Full citations, equations, ablation data, and per-module implementation audit: **[docs/papers/science.md](docs/papers/science.md)** | **[Research post on structured context assembly](docs/research-post-context-assembly.md)**
+**45 papers total.** Every algorithm, constant, and threshold traces to a published source. Full citations, equations, ablation data, and per-module implementation audit: **[docs/papers/science.md](docs/papers/science.md)** | **[Thermodynamic memory paper (PDF, 30 pages)](docs/arxiv-thermodynamic/main.pdf)** | **[Structured context assembly paper (PDF, 37 pages)](docs/arxiv-context-assembly/main.pdf)** | **[Research post on structured context assembly](docs/research-post-context-assembly.md)**
 
 ---
 
@@ -391,19 +393,38 @@ Every benchmark headline number above is backed by a per-mechanism ablation camp
 - **LoCoMo, 14 rows, n=1986 (pre-plasticity-fix bytes)** — `tasks/e1-v3-locomo-results.md`. Two-baseline (NO_CONSOLIDATION / WITH_CONSOLIDATION) design; empirical resolution of the architectural-mismatch hypothesis (RECONSOLIDATION ΔMRR = +0.0076, ADAPTIVE_DECAY ΔMRR = -0.0163).
 - **LoCoMo, 14 rows, n=1986 (post-plasticity-fix bytes)** — `tasks/e1-v3-locomo-results-post-fix.md`. Re-run on commit `2f45bcb` (descendant of plasticity result-shape fix `5f737fe`); cadence-fix anchor agreement re-validated identically (ΔvsNO = +0.0014); two consolidation-only rows (HOMEOSTATIC_PLASTICITY, SCHEMA_ENGINE) recover positive contributions previously masked by the contract bug.
 
-Total: 45 per-mechanism evidence rows. The full paper, including the §6.3 per-mechanism evidence section and §6.3.4.1 plasticity-fix re-run subsection, is at `docs/arxiv-thermodynamic/main.pdf`.
+Total: 45 per-mechanism evidence rows across 26 enum mechanisms (17 read-path + 9 consolidation-only routed to LoCoMo). The full thermodynamic memory paper, including §6.3 per-mechanism evidence and §6.3.4.1 plasticity-fix re-run subsection, is at `docs/arxiv-thermodynamic/main.pdf` (30 pages, all 45 citations resolved). The companion structured context assembly paper is at `docs/arxiv-context-assembly/main.pdf` (37 pages).
 
 ## License
 
 MIT
 
 ## Citation
 
+If you reference the system, the paper PDFs on `main` are the canonical artefacts (arXiv IDs forthcoming, endorsement in progress):
+
 ```bibtex
 @software{cortex2026,
   title={Cortex: Persistent Memory for Claude Code},
   author={Deust, Clement},
   year={2026},
   url={https://github.com/cdeust/Cortex}
 }
+
+@unpublished{deust2026thermodynamic,
+  title={Thermodynamic Memory for Conversational Agents:
+         A Per-Mechanism Ablation Study on LongMemEval and LoCoMo},
+  author={Deust, Clement},
+  year={2026},
+  note={arXiv ID forthcoming, endorsement in progress},
+  url={https://github.com/cdeust/Cortex/blob/main/docs/arxiv-thermodynamic/main.pdf}
+}
+
+@unpublished{deust2026context,
+  title={Structured Context Assembly for Long-Horizon Conversational Memory},
+  author={Deust, Clement},
+  year={2026},
+  note={arXiv ID forthcoming, endorsement in progress},
+  url={https://github.com/cdeust/Cortex/blob/main/docs/arxiv-context-assembly/main.pdf}
+}
 ```