docs(verif): E1 v3 LoCoMo post-plasticity-fix integration — paper §6.3 third pass + README/CLAUDE.md sync

LoCoMo 14-row sweep on bytes including plasticity result-shape fix
(commit 5f737fe). BASELINE_NO=0.8279, BASELINE_WITH=0.8265 (cadence
fix re-validated at full n=1986). Post-fix sign flips: HOMEOSTATIC
-0.0025→+0.0017; SCHEMA_ENGINE -0.0004→+0.0017. Architectural-mismatch
hypothesis holds (RECONSOLIDATION +0.0091 LoCoMo vs +0.0000 LME-S;
ADAPTIVE_DECAY -0.0163 amplified 11x).

Files:
- tasks/e1-v3-locomo-results-post-fix.md (new writeup)
- docs/papers/thermodynamic-memory-vs-flat-importance.md (§6.3 third pass)
- docs/arxiv-thermodynamic/main.tex (LaTeX mirror)
- docs/arxiv-thermodynamic/main.pdf (recompiled, 27 pages)
- README.md (LongMemEval + LoCoMo tables + Verification section)
- CLAUDE.md (scores table)

Closes task #58, task #59. E1 v3 verification campaign complete:
17 LME-S rows + 14 LoCoMo rows + 14 LoCoMo post-fix rows = 45 row
entries documenting per-mechanism evidence on the appropriate
benchmark for each mechanism's mechanism-of-action.

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

Author: cdeust

CLAUDE.md

@@ -341,10 +341,10 @@ python3 benchmarks/episodic/run_benchmark.py --events 20           # Episodic Me
 **Current benchmark scores (clean DB, April 2026):**
 | Benchmark | Cortex | Best in paper |
 |---|---|---|
-| LongMemEval R@10 | **97.8%** | 78.4% |
-| LongMemEval MRR | **0.882** | -- |
-| LoCoMo R@10 | **92.6%** | -- |
-| LoCoMo MRR | **0.794** | -- |
+| LongMemEval R@10 | **98.4%** | 78.4% |
+| LongMemEval MRR | **0.9124** | -- |
+| LoCoMo R@10 | **94.2%** | -- |
+| LoCoMo MRR | **0.8278** | 0.794 |
 | BEAM Overall | **0.591** | 0.329 |
 
 ## Research-Driven Improvement Workflow

README.md

@@ -132,17 +132,17 @@ LongMemEval (Wu et al., ICLR 2025): 500 human-curated questions embedded in ~40
 
 | | Cortex | What it means |
 |---|---|---|
-| Recall@10 | **97.8%** | The right memory shows up in the top 10 results for nearly every question |
-| MRR | **0.882** | The correct answer is usually the first or second result |
+| Recall@10 | **98.4%** | The right memory shows up in the top 10 results for nearly every question |
+| MRR | **0.9124** | The correct answer is usually the first or second result |
 
 | Category | MRR | R@10 | Why this score |
 |---|---|---|---|
-| Single-session (assistant) | 0.982 | 100.0% | Verbatim assistant responses are easy to match |
-| Multi-session reasoning | 0.936 | 99.2% | Entity graph connects evidence across sessions |
-| Knowledge updates | 0.921 | 100.0% | Heat decay naturally surfaces the newest version of a fact |
-| Temporal reasoning | 0.857 | 97.7% | Time anchors embedded directly in memory content |
-| Single-session (user) | 0.806 | 94.3% | User phrasing varies more than assistant responses |
-| Single-session (preference) | 0.641 | 90.0% | Preferences are implicit — harder to retrieve by keyword |
+| Single-session (assistant) | 1.000 | 100.0% | Verbatim assistant responses are easy to match |
+| Multi-session reasoning | 0.962 | 100.0% | Entity graph connects evidence across sessions |
+| Knowledge updates | 0.925 | 100.0% | Heat decay naturally surfaces the newest version of a fact |
+| Temporal reasoning | 0.926 | 98.5% | Time anchors embedded directly in memory content |
+| Single-session (user) | 0.814 | 94.3% | User phrasing varies more than assistant responses |
+| Single-session (preference) | 0.668 | 93.3% | Preferences are implicit — harder to retrieve by keyword |
 
 Knowledge updates scored highest because heat-based decay naturally pushes newer information above older versions of the same fact. This wasn't designed for the benchmark. It's just how the thermodynamic model works.
 
@@ -152,16 +152,16 @@ LoCoMo (Maharana et al., ACL 2024): 1,986 questions across 10 conversations, inc
 
 | | Cortex | What it means |
 |---|---|---|
-| Recall@10 | **92.6%** | Right memory in top 10 over 9 times out of 10 |
-| MRR | **0.794** | Correct answer is typically the first result |
+| 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 |
 
 | Category | MRR | R@10 | Why this score |
 |---|---|---|---|
-| Adversarial | 0.855 | 93.9% | Trick questions can't fool five fused signals |
-| Open-domain | 0.835 | 95.0% | Broad questions benefit from multi-signal coverage |
-| Multi-hop | 0.760 | 88.8% | Entity graph connects evidence across turns |
-| Single-hop | 0.700 | 92.9% | Direct factual questions — strong but room to improve |
-| Temporal | 0.539 | 77.2% | "When did X happen?" is the hardest category — needs better time-series matching |
+| Adversarial | 0.881 | 96.0% | Trick questions can't fool five fused signals |
+| Open-domain | 0.875 | 96.9% | Broad questions benefit from multi-signal coverage |
+| Multi-hop | 0.779 | 90.3% | Entity graph connects evidence across turns |
+| Single-hop | 0.741 | 94.0% | Direct factual questions — strong but room to improve |
+| Temporal | 0.577 | 78.3% | "When did X happen?" is the hardest category — needs better time-series matching |
 
 No LLM at query time. No API calls. Just a 22MB embedding model, PostgreSQL with pgvector, and neuroscience algorithms doing the heavy lifting. Five retrieval signals fused server-side (vector similarity, full-text search, trigram matching, thermodynamic heat, recency), then reranked by a cross-encoder.
 
@@ -383,6 +383,16 @@ ruff format --check .     # Format
 
 ---
 
+## Verification
+
+Every benchmark headline number above is backed by a per-mechanism ablation campaign on the appropriate benchmark for each mechanism's mechanism-of-action. The campaign comprises three artefact sets at full n on a single-seed protocol with code SHAs, dirty flags, manifests, and per-row JSON outputs preserved alongside the writeups:
+
+- **LongMemEval-S, 17 rows, n=500** — `tasks/e1-v3-results.md`. Per-mechanism deltas across the integrated stack at the calibrated equilibrium; category-specialization analysis.
+- **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`.
+
 ## License
 
 MIT