# Changelog

## Score

- CAI 43.8 → 42.1 (-1.7)
- Rubric changed (rubric-2026.11.2 → rubric-2026.11.4) — scores are not directly comparable.

## Lenses

- Code Health 53.3 → 52.2 (-1.1)
- Architecture 95.3 → 95.3 (0.0)
- Maturity 88.8 → 89.2 (+0.4)
- Readiness 27.7 → 22.0 (-5.7)
- Security 88.6 → 88.6 (0.0)
- Accessibility 39.4 → 43.2 (+3.8)
- Performance 100.0 → 100.0 (0.0)

## Resolved (30)

- Classify ADR enforcement: Candidate-pool union fusion (FTS ∪ dense ∪ MV → reranker) — workload-dependent, not a default (eval/decisions/union-fusion-2026-08-20.md)
- Classify ADR enforcement: Code-review fix-set: measured impact (arm M / "fixed") — 2026-08-17 (eval/decisions/fixset-impact-2026-08-17.md)
- Classify ADR enforcement: Completeness-clause A/B (arm J) — 2026-08-15 (eval/decisions/completeness-clause-ab-2026-08-15.md)
- Classify ADR enforcement: Component ablation study — 2026-08-18 (eval/decisions/component-ablation-2026-08-18.md)
- Classify ADR enforcement: Cross-bench validation: are the RFC-tuned changes general? (2026-08-15) (eval/decisions/cross-bench-validation-2026-08-15.md)
- Classify ADR enforcement: Experiments: resolve-only decomposition + LateOn MaxSim reranker — 2026-08-19 (eval/decisions/experiments-resolveonly-maxsim-2026-08-19.md)
- Classify ADR enforcement: FTS index on late-chunk tables (arm K) — 2026-08-15 (eval/decisions/ftslc-index-fix-2026-08-15.md)
- Classify ADR enforcement: GLM-OCR feasibility spike — Apple Silicon (roadmap Phase 1.3) (eval/decisions/glm-ocr-spike.md)
- Classify ADR enforcement: Multi-turn decomposer: answer-visibility fix (arms m0–m1d, L) — 2026-08-16 (eval/decisions/multiturn-decomposer-2026-08-16.md)
- Classify ADR enforcement: Multi-vector (late-interaction) first-stage retrieval — NOT adopted (eval/decisions/multivector-retrieval-2026-08-19.md)
- Classify ADR enforcement: Paraphrase-robustness study: does multi-vector help once queries stop matching document wording? (eval/decisions/paraphrase-robustness-2026-08-20.md)
- Classify ADR enforcement: Phase 1.1 decision — reranker A/B, measured 2026-08-09 (eval/decisions/reranker.md)
- Classify ADR enforcement: Phase 1.2 — embedder audit + A/B — measured 2026-08-09 (eval/decisions/embedder.md)
- Classify ADR enforcement: Phase 2 pipeline-shape items — 2.5, 2.1, 2.2, 2.4 (eval/decisions/phase2-pipeline.md)
- Classify ADR enforcement: Phase 2.3 — cheapen gateway routing — shipped 2026-08-09 (eval/decisions/phase2-gateway.md)
- Classify ADR enforcement: Phase 4 item 4.2 — making the cross-reference hop measurable, and the first A/B (eval/decisions/phase4-eval-final-metric.md)
- Classify ADR enforcement: Phase 4 items 4.1 and 4.2 — end-to-end ANSWER-QUALITY A/Bs (eval/decisions/phase4-answer-quality.md)
- Classify ADR enforcement: Phase 4 items 4.2 and 4.3 — cross-reference hop and overview prefilter (eval/decisions/phase4-crossref-prefilter.md)
- Classify ADR enforcement: Phase 4 items 4.2 and 4.3 — the retrieval benchmark matrix, on rebuilt indexes (eval/decisions/phase4-retrieval-benchmarks.md)
- Classify ADR enforcement: Phase 4 items 4.4 and 4.6 — metadata filter DSL and ephemeral "ask a folder" (eval/decisions/phase4-filters-askfolder.md)
- …and 10 more

## New (46)

- Critical CVE: [GHSA redacted] (requirements-docker.txt)
- Critical CVE: [GHSA redacted] (requirements.txt)
- Edited copy of a member (10 corresponding lines) (backend/database.py)
- Edited copy of a member (11 corresponding lines) (rag_system/agent/escalation.py)
- Edited copy of a member (12 corresponding lines) (backend/database.py)
- Edited copy of a member (14 corresponding lines) (backend/database.py)
- Edited copy of a member (2 corresponding lines) (eval/build_goldset.py)
- Edited copy of a member (39 corresponding lines) (rag_system/utils/ollama_client.py)
- Edited copy of a member (4 corresponding lines) (rag_system/utils/ollama_client.py)
- Edited copy of a member (5 corresponding lines) (rag_system/ingestion/chunking.py)
- Edited copy of a member (9 corresponding lines) (backend/server.py)
- High CVE: [GHSA redacted] (package-lock.json)
- High CVE: [GHSA redacted] (requirements-docker.txt)
- High CVE: [GHSA redacted] (requirements.txt)
- Low coverage: backend/database.py (backend/database.py)
- Low coverage: backend/ollama_client.py (backend/ollama_client.py)
- Low coverage: backend/server.py (backend/server.py)
- Low coverage: create_index_script.py (create_index_script.py)
- Low coverage: rag_system/agent/escalation.py (rag_system/agent/escalation.py)
- Low coverage: rag_system/agent/loop.py (rag_system/agent/loop.py)
- …and 26 more

## Architecture

- Unchanged — 0 containers · 1 contexts · 0 edges
