A/B eval (eval_retrieval.py, 86-query gold-set) showed the 0.5 default was
mis-tuned: the image side was too heavy and dragged precedent_library recall
0.971 -> 0.885. Sweep 0.5..0.75 — at 0.65 multimodal beats text-only on every
overall metric AND every corpus (R@5 0.994 vs 0.989, nDCG@5 0.960 vs 0.944,
MRR 0.954 vs 0.936). Dafna approved.
- MULTIMODAL_TEXT_WEIGHT=0.65 set in Coolify (legal-ai, runtime) + redeploy.
- baseline.json updated to the 0.65 config (future regression reference).
- #15 done (premise was stale — multimodal already default on 110 docs; the
win was tuning the weight, not the backfill).
- #80 opened: the costly 140-doc legacy backfill is deferred until a targeted
image-answer gold-set proves the table/image value prop (untested here).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The gold-set was 77 known-item probes (query=case_name). Added 9 chair-approved
SEMANTIC queries (S1–S9) — a real legal question per row, relevant = the
precedents that should surface (drawn from subject_tags, chair-confirmed). These
test what matters: does retrieval answer a legal issue, not just find a case by
name. source='chair' (preserved across re-bootstrap). practice_area left empty
so the filter never excludes a cross-tagged precedent (s.197 rulings sit under
betterment_levy).
Baseline now 86 queries. Finding from the 9 semantic queries: MRR ≈ 1.0 — the
system surfaces a lead relevant precedent at rank 1 for nearly every question —
but R@10 ranges 0.5–1.0: for broad questions with many co-relevant precedents
(e.g. נטרול תמ"א 38 = 5 relevant → R@10 0.60; שמאי מכריע = 2 → 0.50) some
co-relevant rulings miss the top-10. Lead-precedent retrieval is strong;
exhaustive multi-precedent recall is the gap.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Chair review of the FU-5 gold-set surfaced one internal_committee record whose
case_name was a code ("ARAR-24-9002") rather than a real name. Per the chair's
citation (ערר 9002/24 קרקעות ירושלים 2 בע"מ נ' הוועדה המקומית ירושלים, נבו
13.8.2025, a s.197 compensation appeal), case_name corrected in the DB to
"קרקעות ירושלים 2" (case_number 9002-24 and citation_formatted were already
correct; only 1 such code-named record exists corpus-wide). Re-bootstrapped the
gold-set (the known-item query is now the real name) and refreshed baseline
(aggregate unchanged — the case retrieves identically under the corrected name).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Covers GAP-11 (INV-RET4/G8) and GAP-14 (INV-QA1/G10). Retrieval quality was
never measured (only telemetry observation) and the halacha review backlog was
invisible (the 10/19 gap was found by accident).
Unit B — backlog visibility (pure code, container):
- metrics.halacha_backlog(conn) → {pending_review, approved, rejected, published,
total, oldest_pending_at}; surfaced in metrics.get_dashboard() (get_metrics MCP
tool) and /api/system/diagnostics. Live count revealed 178 pending / 1552 total,
oldest from 2026-05-03 — previously invisible.
Unit A — retrieval eval harness (host-side scripts):
- scripts/eval_gold_bootstrap.py — seeds data/eval/gold-set.jsonl. Two sources:
citations (cited==relevant via search_relevance_feedback — empty until decisions
cite precedents) and known_item (query=case_name → relevant=self; a real
citation-free signal, the methodology #52 checked by hand). Idempotent; preserves
source='chair' rows.
- scripts/eval_retrieval.py — runs the production retrieval path (search_library /
search_internal) over the gold-set; computes precision@k, recall@k, MRR, nDCG@k
(k=5,10); aggregates overall + per-corpus + per-practice_area; writes a report and
a delta vs committed baseline.json (which records the retrieval_config it reflects).
--self-test unit-checks the metric math offline.
Gold-set strategy = hybrid (chair decision): bootstrap + chair review. The citation
source is empty today (0 cited precedents in decisions), so the seed is known-item
(77 queries: 54 internal_decisions + 23 precedent_library). The gold-set is
PROVISIONAL until Dafna reviews it (the domain chair-gate).
Baseline (production config: multimodal+rerank on): R@10=0.987, MRR=0.837,
nDCG@10=0.872. Finding: MULTIMODAL_ENABLED=true slightly lowers known-item recall
(image-page results displace exact name matches) — relevant to #15. precedent_library
weaker than internal (R@10 0.957 vs 1.0) — one external precedent unfindable by name.
"CI gate" realized as discipline (re-runnable harness + committed baseline + run
before/after any retrieval-layer change) — retrieval needs prod DB + Voyage, no CI
runner has that access.
Spec: docs/superpowers/specs/2026-05-31-fu5-eval-harness-design.md
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
