What changed — rapidmeta-finerenone data-integrity audit

Latest sweep — 2026-05-25 → 2026-05-26 (4 commits)

Performance + encoding-correctness pass. Every change went through the existing CI pipeline (idempotency + sentinel + puppeteer smoke).

Performance

• Inline RapidMeta engine minified in 1,524 review pages — terser pass (mangle disabled to preserve the global RapidMeta API) cut each engine block from ~970 KB to ~585 KB. Total: 1.46 GB → 889 MB across the portfolio; 569 MB saved (39.1%). Lighthouse "total byte weight" audit moves from CRITICAL to MEDIUM on every FULL_REVIEW. commit 88e7b07e1
• Auto-minify hook in clone_dashboard.py — freshly cloned pages now go through terser on write, so the engine never re-inflates between bulk-clone runs.
• Sentinel R8 (un-minified engine guard) — pre-push hook now BLOCKs any *_REVIEW.html whose largest inline <script> body is ≥ 800 KB (post-minification typical is ~585 KB; anything larger means the engine was reintroduced unminified).

Encoding correctness

• 1,321 dashboards repaired from cp1252-roundtrip em-dash mojibake (1,306 in repo root + 15 mirrors in e156-submission/assets/). The R-Validation comment "RapidMeta v11.0 R-Validation (DL + HKSJ + Bayesian) — HR" had its em-dash mis-decoded as â€" by a one-time editor save; byte-level repair restored UTF-8 e2 80 94.
• Sentinel R7 (mojibake guard) — BLOCKs any reintroduction of the 3-char mojibake sequence â€". Verified to fire on synthetic input and stay silent on the cleaned corpus.
• Generator-level safety verified: clone_dashboard.py reads/writes encoding='utf-8' throughout and uses DUPILUMAB_COPD_REVIEW.html (now mojibake-free) as its base template. Future clones inherit the clean bytes.

Idempotency + tests

• 17 idempotency tests total (was 9 last session) — every fix script in scripts/ verified to be a no-op on its own output. New tests cover: PI propagation, the 4 a11y contrast fixers, PICO input labels, heading-order lite, Plotly externalisation, em-dash mojibake repair, and the terser minifier itself.
• 8 Sentinel rules (was 6 last session) — R7 mojibake + R8 un-minified engine.

Honest deferrals

• FULL_REVIEW engine externalisation to vendor/ would yield a bigger Lighthouse win (browser caches one engine across many pages) but the engine has 138 hardcoded drug/slug references across 95 lines interleaved inside JS string literals, regex literals, and Python-code-in-template-literals blocks. A clean externalisation requires an AST-aware refactor (acorn/esprima walking only safe StringLiteral nodes) — not feasible in one session. Minification is the safest one-session win on the same byte cost.
• One file untracked: HIV_NEXTGEN_ART_REVIEW.html has a pre-existing JS syntax bug (unterminated string at engine line 37638; propagate_robme_engine.py appears to have injected raw JS where escaped string concatenation was needed). Excluded via .gitignore so it doesn't poison git add *.html; Sentinel R8 surfaces it as a persistent warning until repaired.

Prior sweep — 2026-05-24 → 2026-05-25 (16 commits)

Two-day sweep across the whole portfolio. Every change shipped behind R-metafor parity validation and a 3-job CI pipeline (idempotency + sentinel + puppeteer smoke).

Reliability & correctness

• Page-breaking JS fixed across 1,123 FULL_REVIEW dashboards — Python None literals replaced with JS null (12,939 occurrences). Pages that previously loaded as static shells now run the full RapidMeta interactive engine.
• 6 individual JS bugs caught by puppeteer smoke — Plotly SVG-title injection (4 files, unescaped " inside template string), unresolved git merge-conflict markers (1 file), orphan return after corrupted method (1 file). All repaired.
• 5 catastrophic R-validation sidecars regenerated from curated publishedHR via metafor REML+HKSJ: COPD_TRIPLE, FGFR_INHIBITORS_SOLID, HEPATITIS_HCV_DAA, HPV_DOSE_REDUCTION, MDRTB_BPAL. Catastrophic parity failures: 5 → 0.
• Sentinel pre-push hook + GH Actions CI (3 jobs: idempotency tests, sentinel 6-rule static scan, puppeteer smoke) — every regression of the bugs we fixed this session is now blocked at commit time AND verified at runtime.

Statistical methodology

• Prediction interval propagated to Cochrane v6.5 t_{k-1} across 213 dashboards (was IntHout 2016 t_{k-2}). Matches metafor predict() default. commit ed61bbf1c
• R-metafor parity gate — every R-validation sidecar verified against a fresh metafor REML+HKSJ pool. 148 pass at 1×10⁻⁶ strict; portfolio is now formally parity-checkable. methodology notes
• AACT param_type-aware extraction — 720 FULL_REVIEW files re-extracted (1,083 trial cells). Percentages and continuous outcomes now distinguished from raw counts; previously the value-vs-N heuristic mislabelled ~30% of trial data.
• Continuous outcomes activated — 814 NCT-level effects (HR 335, MD 258, OR 137, RD 72, RR 12) from AACT outcome_analyses.txt + 545 more from per-arm Mean+SD with relaxed dispersion-type matching. 1,359 trials' MD/HR/OR/RR/RD now poolable that previously had null event counts.
• RoB-2 derived from AACT designs.txt — 2,615 trial RoB arrays across 1,306 files. Replaces the misleading ['low','low','low','low','low'] stub with a defensible D1–D5 heuristic from allocation + masking + assessor-blinding fields.
• PMIDs corrected — 2,019 + 1,439 trial PMID links fixed using AACT study_references with "smallest RESULT/DERIVED PMID" heuristic (the trial's original primary publication, not extensions).

Content & usability

• 1,093 per-topic protocol files generated from each FULL_REVIEW's structured PICO data — header badges now resolve to a real protocol, not a fallback redirect. Each carries an "auto-generated, NOT a pre-registered PROSPERO record" banner.
• Schema.org MedicalScholarlyArticle JSON-LD on every page (2,178 total) — LLM & search-engine surface improved.
• 28,897 aria-labels added to icon-only buttons + 2,178 skip-to-main-content links — every download icon, toggle, and toolbar action now reads cleanly in screen readers.
• Audit-table bands recomputed with continuous-outcome-aware logic. Trustworthy: 866 → 2,005; Quarantined: 10 → 0.
• 1,507 meta description tags + sitemap.xml + robots.txt for SEO/LLM discoverability.
• AUTO_REVIEW lite pages got a banner CTA to their full-interactive twin + per-trial PubMed abstract hydration with embedded NCT→PMID map.

Internal docs & tests

• ~/.claude/rules/advanced-stats.md updated — PI formula resolved to t_{k-1} per Cochrane v6.5 (was IntHout t_{k-2}).
• scripts/tests/test_idempotency.py — 9 tests verifying every fix_*.py/apply_*.py is idempotent (prevents the compounding regression class that caused cE 524 → 204 → 80 → 31 in an earlier fix).
• scripts/_js_parse_gate.py — build-time Node parse-check wired into bulk_clone + retro fixers; rolls back any post-rewrite JS that fails to parse.

Pre-session state (preserved for historical reference)

Post-session: see audit_table.html for the current banded counts (Trustworthy 2,005 / LowConcern 82 / ManualReview 67 / Quarantined 0 as of 2026-05-25).

Per-trial identity-confidence score (R33 — 1,655 trials)

Composite of 25%·NCT-in-AACT + 20%·drug-in-AACT-intvs + 15%·PMID-present + 10%·year + 10%·name + 20%·evidence[]-non-empty.

Audit rounds (chronological)

Ground-truth sources

• AACT 2026-04-12 — Clinical Trials Transformation Initiative pipe-delimited snapshot (studies, conditions, interventions, baseline_counts, outcome_measurements, design_outcomes, id_information, primary_completion_date)
• PubMed E-utilities — NCBI metadata (1,190 cached abstracts) + idconv API (PMID↔PMC mapping)
• PMC OAI full-text — NCBI PubMed Central full-text XML (60 papers retrieved, 12 evidence entries injected at ≥3 number matches)
• Retraction Watch (Crossref/GitLab) — 70,147 retractions cross-checked; 0 retracted papers in corpus
• Published landmark trials — NEJM/Lancet primary publications (16 trials, 34 occurrences, 97% byte-exact)
• Published reference MAs — 20 published meta-analyses, REML+HKSJ pool reproduction (93% within Cochrane floor)
• Multi-agent consensus — 3-lens ≥2-of-3 consensus (R1-R4) + 8-agent blinded TruthCert HMAC (157 reviews)
• Internal cross-consistency — direction-inconsistent, copy-paste arms, NCT-era, PMID-cross-NCT, baseline-N mismatch, CI-outside-HR, year-inconsistent-across-reviews, MeSH-anchored topic matching

Files

• FINAL_INTEGRITY_REPORT_V2.md — 232-line consolidated methodology
• fabrication_risk_table.csv — per-review classifier scores
• truthcert/ — 409 HMAC-SHA256 signed per-review manifests
• scripts/ — every round's auto-fix script (replayable)