feat(learning): voice-fingerprint corpus-refresh proposal generator (#161, prose half)
The voice-fingerprint (docs/daphna-voice-fingerprint.md) is the PRIMARY abstract style channel the writer consumes (07-learning §0.2 channel A), but it was hand-authored and never regenerated from the growing corpus — the "voice" the writer adapts stayed frozen while finals accumulated. PR #345 unfroze the structural half (corpus-measured ratios → writer); this is the prose half. scripts/regenerate_voice_fingerprint.py reads the corpus (style_corpus stats + a bounded block-level style_exemplars sample + measured section ratios) and asks Opus (local) to PROPOSE refinements/additions to the fingerprint — STYLE / METHOD / VOICE / LEXICON only, grounded in corpus evidence, never case substance (INV-LRN5), preserving curated insights rather than replacing them. CHAIR-GATED, HARD (INV-LRN1): the prose profile is voice-knowledge, not a low-risk style rule, so it does NOT auto-flow. The script writes a PROPOSAL to data/curator-proposals/ (the existing curator-proposal surface) and NEVER overwrites the live fingerprint; the chair reviews and hand-commits accepted parts — the same manual gate the fingerprint already uses. Runs on the host (claude CLI). Reuses claude_session + measure_corpus_ratios + the curator-proposals dir (G2 — no parallel synthesis or gate). SCRIPTS.md + spec §0.2 updated. Invariants: INV-LRN1 (hard chair gate for voice-knowledge prose), INV-LRN5 (style-only synthesis), G2 (reuses existing synthesis inputs + proposal surface). Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
This commit is contained in:
@@ -101,6 +101,7 @@
|
||||
| `notify.py` | python | שליחת מייל התראה מסוכנים via SMTP (Gmail) | נקרא ע"י סוכנים |
|
||||
| `bidi_table.py` | python | יצירת טבלאות box-drawing עם תמיכה ב-BiDi (עברית+אנגלית) | ספריית עזר |
|
||||
| `convert_decision_template.py` | python | המרת `data/training/טיוטת החלטה.dotx` → `skills/docx/decision_template.docx` לטעינה ב-python-docx | להריץ כשמתעדכנת התבנית |
|
||||
| `regenerate_voice_fingerprint.py` | python | מסנתז מהקורפוס (style_corpus + דוגמאות + יחסים) **הצעה** לעדכון פרופיל-הקול (Opus מקומי). שער-יו"ר קשיח (INV-LRN1): כותב ל-`data/curator-proposals/`, לעולם לא דורס את `docs/daphna-voice-fingerprint.md` — היו"ר סוקרת ומקבעת ידנית. החצי-הפרוזה של #161 (החצי-המבני=ratios ב-PR#345). רץ על המארח (claude CLI). STYLE/METHOD בלבד (INV-LRN5) | ידני / לתזמן תקופתית |
|
||||
| `deploy-track-changes.sh` | bash | סנכרון skills CMP↔CMPA + בדיקות + הנחיות deploy לארכיטקטורת Track Changes | ידני |
|
||||
|
||||
## חד פעמי
|
||||
|
||||
113
scripts/regenerate_voice_fingerprint.py
Executable file
113
scripts/regenerate_voice_fingerprint.py
Executable file
@@ -0,0 +1,113 @@
|
||||
#!/usr/bin/env python3
|
||||
"""Propose a corpus-refreshed update to Dafna's voice-fingerprint (TaskMaster #161).
|
||||
|
||||
The voice-fingerprint (docs/daphna-voice-fingerprint.md) is the PRIMARY abstract
|
||||
style channel the writer consumes (07-learning §0.2, channel A). It was hand-authored
|
||||
and never regenerated from the growing corpus — so the "voice" the writer adapts has
|
||||
been frozen while the corpus grew. This script reads the corpus (style_corpus stats +
|
||||
block-level style_exemplars sample + measured section ratios) and asks Opus (local)
|
||||
to PROPOSE refinements/additions to the fingerprint — STYLE / METHOD / VOICE / LEXICON
|
||||
ONLY, never case substance (INV-LRN5).
|
||||
|
||||
CHAIR-GATED (INV-LRN1, HARD gate — the prose profile is voice-knowledge, not a
|
||||
low-risk style rule, so it does NOT auto-flow): the proposal is written to
|
||||
data/curator-proposals/ and NEVER overwrites the live fingerprint. The chair reviews
|
||||
it (in /training, with the other curator proposals) and hand-commits the parts she
|
||||
accepts to docs/daphna-voice-fingerprint.md — the same manual gate the fingerprint
|
||||
already uses. This is the prose half of the voice-profile refresh; the structural
|
||||
half (corpus-measured ratios → writer) shipped in PR #345.
|
||||
|
||||
Runs on the HOST (claude_session needs the local claude CLI). Usage:
|
||||
DOTENV_PATH=/home/chaim/.env DATA_DIR=/home/chaim/legal-ai/data \\
|
||||
/home/chaim/legal-ai/mcp-server/.venv/bin/python scripts/regenerate_voice_fingerprint.py
|
||||
"""
|
||||
from __future__ import annotations
|
||||
|
||||
import asyncio
|
||||
import logging
|
||||
from datetime import datetime, timezone
|
||||
|
||||
from legal_mcp import config
|
||||
from legal_mcp.services import claude_session, db
|
||||
from legal_mcp.services.style_distance import measure_corpus_ratios
|
||||
|
||||
logging.basicConfig(level=logging.INFO, format="%(message)s")
|
||||
log = logging.getLogger("voice_fingerprint_refresh")
|
||||
|
||||
FINGERPRINT = config.DATA_DIR.parent / "docs" / "daphna-voice-fingerprint.md"
|
||||
PROPOSALS = config.DATA_DIR / "curator-proposals"
|
||||
_PER_SECTION = 5 # exemplars per section in the evidence pack
|
||||
_MAX_PARA_CHARS = 700 # cap each exemplar so the prompt stays bounded
|
||||
|
||||
_SYSTEM = """אתה מזקק את טביעת-אצבע-הקול של עו"ד דפנה תמיר — פרופיל-סגנון מופשט (איך היא כותבת: קול, שיטה, מבנה, לקסיקון, אנטי-דפוסים), לא תוכן.
|
||||
חוקים מחייבים:
|
||||
- **סגנון/שיטה בלבד (INV-LRN5):** אסור לכלול מהות משפטית ספציפית — הלכה, עובדה, תקדים, או ניסוח מתיק קונקרטי. רק הכללות-סגנון.
|
||||
- **לחדד ולהרחיב, לא לזרוק:** שמר את התובנות האצורות בפרופיל הנוכחי; הצע תוספות/חידודים מעוגנים בראיות-הקורפוס שלהלן.
|
||||
- **לעגן:** כל קביעה חדשה נשענת על דפוס שחוזר בראיות. אם אין עיגון — אל תמציא ואל תוסיף.
|
||||
- שמר על מבנה-המסמך והסעיפים הקיימים.
|
||||
הפלט: גרסת-fingerprint מוצעת מלאה (Markdown), מוכנה לסקירת-יו"ר — בלי הקדמות או הסברים מסביב."""
|
||||
|
||||
|
||||
async def _gather_evidence() -> tuple[str, str]:
|
||||
pool = await db.get_pool()
|
||||
async with pool.acquire() as conn:
|
||||
n_corpus = await conn.fetchval("SELECT count(*) FROM style_corpus WHERE coalesce(full_text,'') <> ''")
|
||||
n_ex = await conn.fetchval("SELECT count(*) FROM style_exemplars")
|
||||
rows = await conn.fetch(
|
||||
"""SELECT section, decision_number, outcome, paragraph_text FROM (
|
||||
SELECT section, decision_number, outcome, paragraph_text,
|
||||
row_number() OVER (PARTITION BY section
|
||||
ORDER BY decision_number DESC, word_count DESC) AS rn
|
||||
FROM style_exemplars
|
||||
) t WHERE rn <= $1 ORDER BY section, rn""",
|
||||
_PER_SECTION,
|
||||
)
|
||||
ratios = await measure_corpus_ratios()
|
||||
parts = [f"קורפוס: {n_corpus} החלטות, {n_ex} דוגמאות-בלוק.", "", "יחסי-מבנה מדודים מהקורפוס (אחוז-מהסך, לפי תוצאה):"]
|
||||
for outcome, entry in (ratios or {}).items():
|
||||
secs = ", ".join(f"{s} {round(p)}%" for s, p in (entry.get("sections") or {}).items())
|
||||
parts.append(f" - {outcome} (n={entry.get('n', 0)}): {secs}")
|
||||
parts += ["", "דוגמאות-סגנון מייצגות (פסקאות אמיתיות של דפנה — מקור ללמידת-קול, לא להעתקת-מהות):"]
|
||||
for r in rows:
|
||||
para = (r["paragraph_text"] or "").strip()[:_MAX_PARA_CHARS]
|
||||
parts.append(f"\n[{r['section']} · {r['decision_number']} · {r['outcome'] or '—'}]\n{para}")
|
||||
return "\n".join(parts), f"{n_corpus} החלטות / {n_ex} דוגמאות"
|
||||
|
||||
|
||||
async def main() -> int:
|
||||
if not FINGERPRINT.exists():
|
||||
log.error("fingerprint not found: %s", FINGERPRINT)
|
||||
return 1
|
||||
current = FINGERPRINT.read_text(encoding="utf-8")
|
||||
evidence, summary = await _gather_evidence()
|
||||
|
||||
user = (
|
||||
f"## הפרופיל הנוכחי (לחדד ולהרחיב, לא להחליף):\n{current}\n\n"
|
||||
f"## ראיות-קורפוס:\n{evidence}\n\n"
|
||||
"## המשימה:\nהצע גרסת-fingerprint מעודכנת המשלבת חידודים/תוספות-סגנון מעוגנים בראיות לעיל, "
|
||||
"תוך שימור כל התובנות האצורות. סגנון/שיטה בלבד — בלי מהות."
|
||||
)
|
||||
log.info("synthesizing voice-fingerprint proposal from %s (Opus, local)…", summary)
|
||||
proposed = await claude_session.query(
|
||||
user, timeout=claude_session.LONG_TIMEOUT, system=_SYSTEM,
|
||||
model="claude-opus-4-8", tools="",
|
||||
)
|
||||
if not proposed or not proposed.strip():
|
||||
log.error("empty proposal from model — aborting (nothing written)")
|
||||
return 1
|
||||
|
||||
PROPOSALS.mkdir(parents=True, exist_ok=True)
|
||||
ts = datetime.now(timezone.utc).strftime("%Y%m%dT%H%M%SZ")
|
||||
out = PROPOSALS / f"voice-fingerprint-{ts}.md"
|
||||
header = (
|
||||
f"<!-- הצעת-חידוש לטביעת-אצבע-הקול · {ts} · מקור: {summary} -->\n"
|
||||
"<!-- שער-יו\"ר (INV-LRN1): הצעה בלבד. לסקור ולקבע ידנית ל-docs/daphna-voice-fingerprint.md. -->\n\n"
|
||||
)
|
||||
out.write_text(header + proposed, encoding="utf-8")
|
||||
log.info("✓ proposal written (chair-gated, NOT applied): %s", out)
|
||||
log.info(" review in /training → commit accepted parts to %s", FINGERPRINT)
|
||||
return 0
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
raise SystemExit(asyncio.run(main()))
|
||||
Reference in New Issue
Block a user