feat(nevo): backfill leaked preamble + ratio gold-set benchmark (#86)

#86.2 backfill + #86.3 benchmark, plus a #86.1 over-strip fix found en route.

extractor.py
- extract_nevo_ratio(): capture Nevo's מיני-רציו block (editorial holdings
  summary) before it is stripped — a free professional gold-set (#86.3).
- _DECISION_START hardening (#86.2): the merged #86.1 regex over-stripped.
  (a) פסק-דין headers are markdown-wrapped (**פסק  דין**); the old anchor
      required the keyword as the first line char with one separator, so it
      missed the header and matched a citation 32K deep (עמ"נ 50567-07-21,
      losing 45% of the body). Now tolerates leading markdown + 0-3 seps,
      and the final-nun form (דין ן vs דינו נ).
  (b) bare השופט/הנשיא matched CITATIONS ("השופט מ' חשין, פסקה 23"). The
      authoring-judge line ends with a colon; we now require it.

ingest.py
- capture the ratio before stripping and store it on the row (best-effort,
  non-fatal); also strip the text-upload path (was file-only).

db.py
- add case_law.nevo_ratio column (additive); allow it in update_case_law.

scripts/backfill_nevo_preamble.py (#86.2) — dry-run-by-default data migration:
finds historically-leaked rulings, captures ratio→nevo_ratio, rewrites
full_text (+content_hash), reindexes, and FLAGS (never deletes) halachot whose
quote lives in the removed preamble (review_status=pending_review +
nevo_preamble_leak flag). Safety guard: rows with keep%<--min-keep (60) are
excluded from --apply as suspected over-strip. --apply writes backup+manifest
to data/audit/ first. Chair-gated — NOT applied here.

scripts/nevo_ratio_benchmark.py (#86.3) — LLM-as-judge (local claude_session,
zero cost) measures recall/precision/granularity of our halachot vs the Nevo
ratio. Works pre- and post-backfill (reads nevo_ratio, falls back to full_text).

Verified:
- pytest tests/test_nevo_preamble.py — 12 passed (incl. citation/markdown
  over-strip regressions).
- backfill dry-run: 19 leaked rulings, 27 contaminated halachot, all ≥75%
  keep (the 32K over-strip is gone).
- benchmark on בג"ץ 1764/05: recall=0.875 precision=1.0 granularity=1.75x.

Invariants: G1 (normalize at source — strip/capture at ingest, not at read);
no silent swallow (contaminated halachot flagged + reported, not dropped);
data-migration is dry-run-default with backup+manifest, chair-gated.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
This commit is contained in:
2026-06-06 19:45:43 +00:00
parent 12bdec10fa
commit fb51a0e869
7 changed files with 552 additions and 8 deletions

View File

@@ -0,0 +1,173 @@
#!/usr/bin/env python3
"""#86.3 — benchmark halacha-extraction quality against Nevo's מיני-רציו gold-set.
Nevo's editorial מיני-רציו is a free, professionally-written list of a ruling's
holdings. By comparing the halachot WE extracted against it we get an honest,
zero-cost measurement of extraction quality per ruling:
* recall — fraction of Nevo's holdings that our halachot cover
* precision — fraction of our halachot that map to a Nevo holding
* granularity — our_count / nevo_holding_count (over-decomposition signal,
the #81.5 concern: e.g. 14 ours vs 4 Nevo = 3.5x)
The gold-truth ratio is read from ``case_law.nevo_ratio`` (populated by
``backfill_nevo_preamble.py`` / ingest). For rulings not yet backfilled it
falls back to computing the ratio on-the-fly from the stored ``full_text``,
so the harness works before and after the migration.
An LLM-as-judge (local ``claude_session``, zero API cost) does the semantic
mapping — string overlap can't tell "same holding, different words" from a
genuinely new holding. The judge is asked to count, not to rewrite.
Run with the MCP server venv (needs the local ``claude`` CLI):
cd ~/legal-ai/mcp-server
.venv/bin/python ../scripts/nevo_ratio_benchmark.py --case 'בג"ץ 1764/05'
.venv/bin/python ../scripts/nevo_ratio_benchmark.py --all --limit 5
.venv/bin/python ../scripts/nevo_ratio_benchmark.py --all # full corpus
"""
from __future__ import annotations
import argparse
import asyncio
import csv
import json
import sys
from datetime import datetime, timezone
from pathlib import Path
from legal_mcp.services import claude_session, db
from legal_mcp.services.extractor import extract_nevo_ratio
REPO_ROOT = Path(__file__).resolve().parent.parent
AUDIT_DIR = REPO_ROOT / "data" / "audit"
_JUDGE_SYSTEM = (
"אתה בוחן-איכות משפטי. נתונים לך (א) רשימת ההלכות (מיני-רציו) שכתב עורך נבו "
"עבור פסק-דין — אמת-המידה; (ב) רשימת ההלכות שמערכת אוטומטית חילצה מאותו "
"פסק-דין. משימתך: למפות סמנטית בין השתיים (אותו עיקרון משפטי בניסוח שונה = "
"התאמה), ולספור. החזר JSON בלבד, ללא טקסט נוסף."
)
def _judge_prompt(ratio: str, ours: list[str]) -> str:
ours_block = "\n".join(f"{i}. {s}" for i, s in enumerate(ours, 1)) or "(אין)"
return (
f"מיני-רציו של נבו (אמת-מידה):\n{ratio}\n\n"
f"ההלכות שחולצו על-ידי המערכת ({len(ours)}):\n{ours_block}\n\n"
"החזר JSON עם המפתחות:\n"
'{"nevo_holdings": <מספר העקרונות הנפרדים במיני-רציו>,\n'
' "covered": <כמה מעקרונות נבו מכוסים ע"י לפחות הלכה אחת שלנו>,\n'
' "ours_total": <מספר ההלכות שלנו>,\n'
' "ours_mapped": <כמה מההלכות שלנו ממופות לעיקרון נבו כלשהו>,\n'
' "notes": "<עד 2 משפטים: מה הוחמץ / מה עודף>"}'
)
async def _bench_one(row: dict, model: str | None) -> dict:
cn = row["case_number"]
ratio = (row.get("nevo_ratio") or "").strip() or extract_nevo_ratio(row.get("full_text") or "")
result = {"case_number": cn, "nevo_holdings": 0, "covered": 0,
"ours_total": 0, "ours_mapped": 0, "recall": None,
"precision": None, "granularity": None, "notes": "", "error": ""}
if not ratio:
result["error"] = "no mini-ratio"
return result
halachot = await db.list_halachot(case_law_id=row["id"], limit=500)
ours = [h["rule_statement"] for h in halachot
if h.get("review_status") in ("approved", "published", "pending_review")
and (h.get("rule_statement") or "").strip()]
result["ours_total"] = len(ours)
if not ours:
result["error"] = "no extracted halachot"
return result
try:
verdict = await claude_session.query_json(
_judge_prompt(ratio, ours), system=_JUDGE_SYSTEM, model=model, effort="low",
)
except Exception as e: # noqa: BLE001
result["error"] = f"judge failed: {e}"
return result
if not isinstance(verdict, dict):
result["error"] = "judge returned non-dict"
return result
nh = int(verdict.get("nevo_holdings") or 0)
cov = int(verdict.get("covered") or 0)
ot = int(verdict.get("ours_total") or len(ours))
om = int(verdict.get("ours_mapped") or 0)
result.update({
"nevo_holdings": nh, "covered": cov, "ours_total": ot, "ours_mapped": om,
"recall": round(cov / nh, 3) if nh else None,
"precision": round(om / ot, 3) if ot else None,
"granularity": round(ot / nh, 2) if nh else None,
"notes": str(verdict.get("notes") or "")[:300],
})
return result
async def main(args: argparse.Namespace) -> int:
pool = await db.get_pool()
async with pool.acquire() as conn:
if args.case:
rows = await conn.fetch(
"SELECT id, case_number, nevo_ratio, full_text FROM case_law "
"WHERE case_number = $1", args.case,
)
else:
# rulings that have (or can derive) a ratio
rows = await conn.fetch(
"SELECT id, case_number, nevo_ratio, full_text FROM case_law "
"WHERE nevo_ratio <> '' OR full_text LIKE '%מיני-רציו:%' "
"ORDER BY case_number"
)
rows = [dict(r) for r in rows]
if args.limit:
rows = rows[: args.limit]
if not rows:
print("no rulings with a mini-ratio found", flush=True)
return 0
print(f"benchmarking {len(rows)} ruling(s)...", flush=True)
results = []
for i, row in enumerate(rows, 1):
res = await _bench_one(row, args.model)
results.append(res)
if res["error"]:
print(f"[{i}/{len(rows)}] {res['case_number']}: SKIP ({res['error']})", flush=True)
else:
print(f"[{i}/{len(rows)}] {res['case_number']}: "
f"recall={res['recall']} precision={res['precision']} "
f"granularity={res['granularity']}x "
f"(nevo={res['nevo_holdings']}, ours={res['ours_total']})", flush=True)
scored = [r for r in results if r["recall"] is not None]
if scored:
avg = lambda k: round(sum(r[k] for r in scored) / len(scored), 3) # noqa: E731
print(f"\n=== {len(scored)} scored — mean recall={avg('recall')} "
f"precision={avg('precision')} granularity={avg('granularity')}x ===", flush=True)
ts = datetime.now(timezone.utc).strftime("%Y%m%dT%H%M%SZ")
AUDIT_DIR.mkdir(parents=True, exist_ok=True)
out = Path(args.out) if args.out else AUDIT_DIR / f"nevo-ratio-benchmark-{ts}.csv"
with out.open("w", encoding="utf-8", newline="") as f:
w = csv.DictWriter(f, fieldnames=list(results[0].keys()))
w.writeheader()
w.writerows(results)
print(f"report: {out}", flush=True)
return 0
if __name__ == "__main__":
ap = argparse.ArgumentParser(description=__doc__,
formatter_class=argparse.RawDescriptionHelpFormatter)
g = ap.add_mutually_exclusive_group(required=True)
g.add_argument("--case", help="benchmark a single case_number")
g.add_argument("--all", action="store_true", help="benchmark all rulings with a mini-ratio")
ap.add_argument("--limit", type=int, default=None, help="cap the number of rulings")
ap.add_argument("--model", default=None, help="judge model (default: CLI session default)")
ap.add_argument("--out", default=None, help="output CSV path (default: data/audit/)")
args = ap.parse_args()
sys.exit(asyncio.run(main(args)))