עד כה claims.party_name נשאר ריק והצבירה/הכתיבה גזרו את הפרדת-המשיבים
מ-source_document בזמן-ריצה (שדה-מת). עכשיו הוא מאוכלס במקור ונקרא ע"י הצרכנים.
- claims_extractor.extract_and_store_claims: מחתים party_name=source_document
ל-SPLIT_PARTIES (respondent/permit_applicant); single-voice נשאר ''. אותו
כלל SPLIT_PARTIES של הצבירה (G2, import לא-מעגלי).
- SCHEMA_V52: backfill idempotent ל-claims קיימים (רק split + party_name ריק).
- argument_aggregator + block_writer._build_claims_context: קוראים
claims.party_name, עם fallback ל-source_document ל-claims מדור-קודם. התנהגות
זהה (party_name==source_document) — אפס-רגרסיה, אך השדה כעת מקור-האמת (G1).
py_compile + טעינת-שרשרת-ייבוא נקיים. סוגר את הפתוח האחרון של #224.
Invariants: G1 (אכלוס במקור, לא גזירה-בקריאה), G2 (כלל-פיצול יחיד).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
השלמת #224 לבלוק י' (דיון והכרעה): הקשר-הטענות כבר מופרד לפי כתב-תשובה
(PR#401, אותו _build_claims_context), אך ה-prompt של block-yod לא הנחה
להתייחס לכל משיב לגופו. נוסף כלל "צדדים מרובים": התייחס לעמדת כל משיב/כתב-תשובה
בנפרד (משיבות 2-3 מול משיבים 4-6), אל תמזג ל"טענות המשיבים" גורפות, וכשעמדות
מנוגדות — ציין והכרע בנימוק.
שינוי-prompt בלבד; משפיע על ניסוח בלוק י' בהחלטות עם כמה משיבים. py_compile ירוק.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
בלוק ז' של טיוטת-הביניים מיזג את כל המשיבים לסעיף אחד גנרי (או השמיטם) —
כי (א) _build_claims_context קיבץ רק לפי party_role, ו-(ב) ה-prompt נעל
"מבנה קבוע 3 חלקים: עוררים / ועדה מקומית / מבקשי-היתר" — בלי "משיבים" בכלל.
לכן בתיק 1043-02-26 המשיבים 2-3 ו-4-6 לא הופיעו כצדדים מובחנים, אף שהנתונים
(129 טענות-משיבים מפוצלות לפי כתב-תשובה) תקינים ב-claims. הפרדת-המשיבים של
#224 יושמה בצבירה+פאנל בלבד (PR#392) — לא בכותב-הבלוקים.
- `_build_claims_context` מקבץ לפי (party_role, brief) כאשר brief=source_document
ל-SPLIT_PARTIES (respondent/permit_applicant) — **אותו כלל בדיוק כמו הצבירה**
(import SPLIT_PARTIES מ-argument_aggregator, G2). התוצאה: סעיף נפרד לכל
כתב-תשובה ("טענות המשיבים — כתב תשובה משיבות 2 3", "— משיבים 4 6"...).
- prompt בלוק ז': מבנה-צדדים דינמי לפי הכותרות בקונטקסט; הוראה מפורשת לא למזג
משיבים שונים ולשקף עמדות-מנוגדות; יעד-אורך גמיש לכמה כתבי-תשובה.
משפיע גם על בלוק י' (אותו _build_claims_context) — הקשר מופרד. py_compile ירוק.
עבודה על MCP → דורש host reload + הרצה-מחדש כדי לראות בפועל.
Invariants: G2 (נגזרת-הפרדה יחידה משותפת עם הצבירה), G1.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
הצבירה קיבצה את כל המשיבים תחת מפלגה אחת "respondent", כך שכתבי-תשובה
נפרדים (משיבות 2-3 מול משיבים 4-6) עם עמדות שעלולות להיות מנוגדות נבלעו
לרשימה אחת. התגלה בתיק 1043-02-26.
השינוי מפצל את הצדדים הרב-משתתפים (respondent/permit_applicant) לפי
source_document — כל כתב-תשובה משותף = יחידת-ליטיגציה קוהרנטית נפרדת —
ומאחסן את התווית ב-legal_arguments.party_name (מיגרציה V50). עוררים/ועדה
נשארים קבוצה יחידה. בונוס: פיצול ה-129 טענות ל-קריאות-Claude קטנות מתקן
את הכשל-בשקט המקורי (קריאה של 100+ פרופוזיציות החזירה non-JSON והפילה את
כל הצד).
- db.py: SCHEMA_V50_SQL — legal_arguments.party_name (ADD COLUMN IF NOT EXISTS, אידמפוטנטי)
- argument_aggregator: קיבוץ (party, party_name); _build_prompt מקבל שם-כתב; SELECT מחזיר party_name
- UI: PartySection מתת-קבץ לפי כתב-תשובה; טיפוס LegalArgument.party_name
טווח v1: הקיבוץ לפי source_document; אכלוס party_name ב-extractor וזיהוי
עמדות-מנוגדות בבלוק ז/י — המשך ב-#224.
Invariants: G1 (נרמול-במקור — party_name בסכמה, לא תיקון-בקריאה) · G2 (אין
מסלול-צבירה מקביל, אותו pipeline) · §6 (אין בליעה שקטה — הכשל-בשקט תוקן).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
חלוקה פנימית (לנוחות בלבד, לא משנה doc_type) של רשימת מסמכי-התיק:
1. עיקריים (appeal/response/appraisal) · 2. נלווים · 3. פרוטוקול ועדת הערר ·
4. לאחר הדיון. הסיווג נגזר בצד-הלקוח מ-doc_type + שני דגלי-metadata.
שני פקדים חדשים בעורך-התיוג:
- מתג "התקבל אחרי הדיון" → metadata.is_post_hearing (כבר נצרך ע"י בלוק-ח);
גובר על הסיווג-לפי-סוג ומעביר לקבוצה 4.
- בורר "פרוטוקול של: ועדת הערר / ועדה מקומית-מחוזית" → metadata.protocol_scope
(נראה רק כשהסוג protocol); ברירת-מחדל appeal→קבוצה 3, lower→קבוצה 2.
Backend: PATCH /documents/{id} + MCP document_update מקבלים is_post_hearing
ו-protocol_scope, נשמרים ב-metadata JSONB (אותו מסלול כמו appraiser_side).
עיצוב אושר ב-Claude Design (X17): 18k-case-documents-grouped.html.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
הגבול _MAX_PROMPT_CHARS=400_000 היה שומר מלאכותי שתאם את קיר ה-200K-טוקן
של בניית-Opus הסטנדרטית (~400K תווים עברית @~2 תווים/טוקן) — לא גבול-מודל.
Opus 4.8 מציע חלון 1M-טוקן במחיר רגיל; `claude -p` חושף אותו ישירות כ-model-id
`claude-opus-4-8[1m]` (אומת CLI 2.1.196, בלי beta-header). בלוק-י (דיון), שנושא
את התיק המלא כ-source-context (805K/891K תווים), נחסם רק ע"י השומר המיושן.
שינויים:
- block_writer: GENERATION_MODEL_1M + סף אסקלציה _CTX_1M_THRESHOLD_CHARS=350_000.
פרומפט > הסף → `[1m]`; אחרת הבנייה הסטנדרטית (בלוקים קטנים נשארים זולים/מהירים).
- _MAX_PROMPT_CHARS 400_000 → 1_500_000 (~750K טוקנים; מרווח מתחת ל-1M) כשומר-קשיח
סופי גם על בניית-ה-1M; הודעת-השגיאה מפנה לצמצום-source-context.
- model_used (provenance) משקף את הבנייה בפועל (200K/[1m]).
- claude_session: אזהרת-גודל מעודכנת (350K + רמז ל-[1m]).
- ספ 04-analysis-writing §1.3: תיעוד האסקלציה כעידון של אותו pin (לא מסלול מקביל, G2).
Invariants: מקיים G2 (אותו מודל-נעוץ, חלון רחב — לא מסלול מקביל), G1, INV-WR*;
לא בולע שגיאות (שומר-קשיח סופי נשאר). xhigh לבלוק-י ללא שינוי.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Fixes the case where the H1 chip, the pipeline stepper and the manual-changer
dropdown disagreed (e.g. 1043-02-26 on analyst_verified): the status sat between
the canonical 10 and a legacy bucket, so each surface read a different list.
Root fix — ONE authoritative status model + canonicalize the real intermediate
states (chair-approved):
- New mcp-server/.../case_status_model.py — the single registry: ordered
StatusDef list (key/label/description/phase/selectable/terminal/on_enter),
the 5 phases, STATUS_ORDER, phase_of/label_of, and a drift assertion that the
CaseStatus enum matches it. `on_enter` is the seam for a status to *do*
something on entry, declared next to its definition (no dispatcher wired yet).
- models.CaseStatus enum: analyst_verified + research_complete promoted to
first-class canonical statuses (the agents set them — they belong in the
model, not a legacy fallback).
- tools/cases.py: the forward-only STATUS_ORDER guard now derives from the
registry instead of an inline list.
- GET /api/status-model exposes the model so the frontend mirror can be
generated against it.
- web-ui case-status.ts mirrors it: the two intermediates moved from the
legacy map into CASE_STATUSES / PHASES(thinking) / STATUS_LABELS /
STATUS_DESCRIPTIONS; status-badge icons+tones extended. So the chip (status),
stepper (its phase) and dropdown (now includes it) all agree.
Invariants: G2 — collapses the scattered status definitions (enum, inline
STATUS_ORDER, two frontend label maps) onto one backend authority + a documented
frontend mirror; no parallel status list remains. Agent-prompt alignment +
backfill follow in a separate change. py_compile + tsc + eslint clean.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Set block-vav (בלוק ו', רקע/background) generation effort from medium to
low in BLOCK_CONFIG. The #208 empirical effort calibration
(scripts/calibrate_effort.py, report
data/eval/effort-calibration-20260630T155054Z.md, n=7 finals) found that
for block-vav, low effort lands closest to chair Dafna Tamir's finalized
text — winning on BOTH the composite distance metric (low 0.4454 vs
medium 0.4749 vs high 0.5381 vs xhigh 0.5413) AND the cleaner style
signal anti_pattern_total (low 1.57 vs medium 2.29 vs high 4.14). Chair
(chaim) approved adopting low for block-vav.
This is the only block with a clean/confident calibration signal —
block-zayin and block-yod-alef signals were ambiguous/tied and are left
unchanged. No other block effort is touched.
Invariants: G8 (effort empirically calibrated, not by-feel);
claude_session-local-only (block_writer runs via the local MCP/claude_local
generation path). Host-side generation parameter — after merge the host
tree needs `git pull` to take effect.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The "אימות פסיקה" tab reloaded its data on every tab revisit and the first
load took 1.5–3s. Two independent causes, two fixes:
Frontend (web-ui/src/lib/api/citation-verification.ts):
- staleTime 10s → 5min and refetchOnMount: false. Radix Tabs unmounts inactive
TabsContent, so switching back used to trigger a full rebuild once the 10s
window elapsed. Verify mutations still invalidate the query, so edits pull
fresh data.
Backend (mcp-server/.../case_citation_verification.py):
- The per-argument loop ran a Voyage embed + vector search + batched authority
query sequentially (N args × ~300ms waterfall). Extracted to _fetch() and
fanned out with asyncio.gather — order preserved, n_verified counter and
per-argument seen-sets unchanged. First load ~5× faster on multi-argument cases.
Pure read/assembly — no new parallel path (G2), no silent error swallowing
(search/radar still log warnings). Invariants: INV-DIG1/INV-AH unaffected
(still read-only, never cites).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Backend follow-up to INV-UI9 (X6 §"אודיט-נלווה"): make server-side
datetimes explicit so flipping the container to TZ=Asia/Jerusalem cannot
silently skew stored or displayed timestamps. Storage stays UTC; human
display is Israel time.
- audit.py: naive datetime.utcnow() → datetime.now(timezone.utc) for the
TIMESTAMPTZ audit_log.created_at. Under TZ=Asia/Jerusalem the naive UTC
wall-clock would be read as Israel-local and stored ~3h off. (§6: no
silent skew.)
- calibrate_effort.py: _ts() report stamp now Israel-time via
ZoneInfo("Asia/Jerusalem") with a filename-safe "-IL" suffix (script runs
on the UTC host); added a labeled "שעון ישראל" line to the report header
so the chair-facing timestamp is unambiguous. --self-test still ALL PASS.
- research_md.py: display fields made deterministic Israel-tz —
update_chair_position timestamp (datetime.now(IL_TZ)) and parse() mtime
(datetime.fromtimestamp(..., tz=IL_TZ)) — correct regardless of process TZ.
Left intentionally Israel-local-day date.today()/year-cutoff call sites
untouched (drafting/party_claims_summary/block_writer/court_fetch_supreme)
— a Jerusalem committee wants the Israel calendar day; they resolve
correctly once TZ flips.
Invariants: INV-UI9 (store-UTC / display-Israel, server side); §6
(timezone-aware datetimes, no silent ~3h skew).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
WS2b של עיצוב-מחדש זרימת-העבודה. מאפשר ניתוח כתב-הערר לבד ואז איחוד
מסמך-עיקרי שנוסף מאוחר, בלי force-delete גורף, עם diff ליו"ר.
- SCHEMA_V49: documents.claims_extracted_at + claims_extraction_status
(אירוע-חילוץ per-מסמך, לא נגזר מ-doc_type), + אינדקס חלקי
idx_documents_claims_pending. idempotent (ADD COLUMN IF NOT EXISTS).
V48 שמור ל-#357.
- claims_extractor חותם את המסמך אחרי שמירת/אי-מציאת טענות.
- db.primary_docs_not_analyzed (is_primary V47 + claims_extracted_at IS NULL)
+ mark_document_claims_extracted; _row_to_doc חושף claims_analyzed.
- reanalyze_claims (כלי-MCP): snapshot→חילוץ-מאחד רק למסמכים חדשים/לא-נותחו
(store_claims מחליף per-source ⇒ טענות אחרות נשמרות)→aggregate force=True
(מסלול קנוני, מוחק רק legal_arguments)→_impact_diff before↔after ליו"ר.
- workflow_status חושף primary_docs_not_analyzed + next-step.
- ספ: 02-data-model §2ג (דגל לא-נותח), 04-analysis-writing §1.3.
- 5 בדיקות-יחידה ל-_impact_diff/_snapshot (פונקציות טהורות).
Invariants:
- G1 (נרמול-במקור): claims_extracted_at נחתם בנקודת-החילוץ, claims_analyzed
נגזר ממנו — לא תיקון-בקריאה.
- G2 (מסלול קנוני יחיד / merge-not-fork): reanalyze מרחיב את מסלול
claims_extractor+argument_aggregator הקיים, לא forks; האיחוד דרך
store_claims per-source; aggregate force=True מוחק רק נגזר (legal_arguments).
- INV-DM (מודל-נתונים): דגל-אירוע per-מסמך, אינדקס חלקי, מקור-אמת יחיד.
- G10 (שער-אנושי): בדיקת-ההשפעה מוצגת ליו"ר, לא מוחלת אוטומטית.
- INV-TOOL idempotency: SCHEMA_V49 idempotent; ניתוח-מחדש חוזר על מסמך
כבר-נותח הוא refresh נקי (store_methods replace per-source).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
מנגנון-הייצור (claude_session→claude -p) לא העביר model/effort: BLOCK_CONFIG
הגדיר temp+model per-בלוק שהיו מטא-דאטה-מת (block_writer.py:447 העביר רק
prompt/timeout/tools). הייצור רץ על מודל ברירת-המחדל של ה-CLI עם פרומפטים-חופשיים
→ מקור אי-העקביות בטיוטת-הביניים.
תיקון (עקביות-מבנית, WS5):
- נעיצת model="claude-opus-4-8" (GENERATION_MODEL) לכל בלוקי-ה-AI + effort per-בלוק
שמועבר בפועל דרך claude_session.query → `claude -p --model --effort`.
מפת-effort: ה=medium · ו=medium · ז=high · ח=medium · ט=high · יא=high · י=xhigh.
- ניסוח-פתיחה קבוע לבלוק-ה (תמיד "לפנינו ערר…", סוף ל-OR "עניינה של החלטה זו").
- מבנה-פנימי קבוע לבלוק-ה; סדר-בלוקים בייצוא כבר קבוע (_INTERIM_BLOCK_ORDER).
- הסרת שדות temp/model המתים מ-BLOCK_CONFIG; "temp" deprecated (Opus 4.7/4.8 דוחים
temperature→400). decision_blocks.temperature נשמר=0 לתאימות-עמודה בלבד.
- timeout נגזר מ-effort (לא מפיצול sonnet/opus שכבר לא קיים).
- עדכון endpoint reference /api/settings/mcp/blocks + blocks-tab + McpBlock לחשוף effort.
- ספ: docs/block-schema.md §3 (effort מחליף temperature), docs/spec/06-export.md
(דטרמיניזם-מבני של טיוטת-הביניים).
G2 (מקור-אמת יחיד) — אין מסלול-ייצור מקביל: write_interim_draft מתזמר את אותו
write_and_store_block→write_block של ההחלטה המלאה.
claude_session-local-only — הייצור נשאר דרך גשר-ה-CLI, לא SDK/קונטיינר.
INV-EX1 — סדר-ייצוא דטרמיניסטי נשמר; G11 — עקרונות-הכתיבה ל-12 הבלוקים נשמרים.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Golden ratios were measured from the corpus (style_distance.measure_corpus_ratios)
but never reached the writer — it used only the hardcoded STRUCTURE_GUIDANCE text.
So the structural target was frozen, and drafts kept deviating (e.g. discussion
block 10% vs a 40-47% target) with no corrective signal.
_structure_guidance_with_corpus(outcome, practice_area) now appends the
corpus-measured section ratios for the block's outcome to the structure guidance,
at both writer entry points (write_block + get_block_context) via one shared helper
(G2). Because measure_corpus_ratios reads style_corpus — which grows with every
enrolled final (see exemplar-growth, #343) — the writer's structural target now
reflects Dafna's ACTUAL current distribution and improves as the corpus grows.
Deterministic (no LLM); best-effort.
This is the STRUCTURAL half of the voice-profile-refresh (#7). The remaining half —
LLM-regenerating the abstract voice-fingerprint PROSE from the corpus (chair-gated)
— is a distinct larger initiative, deliberately deferred rather than rushed.
Invariants: G1 (the writer gets the live measured target, not a frozen constant),
G2 (one structure-guidance builder shared by both entry points; reuses the existing
measure_corpus_ratios — no parallel measurement), INV-LRN4 (the corpus the loop
grows now feeds the writer's structural target).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The chair-feedback chain was DEAD: 27 feedback rows captured, 0 ever became a
lesson the writer reads (no weekly-analysis job ever existed). The chair's own
corrections — the most authoritative signal of all, and exactly "learn from every
returned draft" — went nowhere.
decision_lessons can't carry them: it's FK-coupled to a style_corpus row (a signed
final), which a case-in-progress doesn't have. So chair STYLE feedback instead
rides the discussion_rules channel that already reaches the writer for all blocks —
the same path /training promote uses.
- db.append_global_rule: the locked read-modify-write append, extracted from web
`_append_methodology_override` into one shared impl (G2). The web function is now
a thin wrapper that seeds defaults; chair-feedback calls it directly.
- record_chair_feedback (MCP tool): a STYLE-category correction (style/wrong_tone/
wrong_structure) with a lesson_extracted flows immediately to discussion_rules.
The chair IS the gate — no separate approval (INV-LRN1 graduated gate). SUBSTANCE
feedback (missing_content/factual_error/other) is case-specific → recorded only.
Invariants: INV-LRN1 (chair-authored style = highest authority, flows; substance
not auto-flowed), G2 (single append impl shared by promote + feedback), INV-LRN5
(style channel only).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The style-exemplar corpus (channel B — the writer's block-level retrieval of
Dafna's real prose) was FROZEN at the one-time seed backfill: new finals were
enrolled into style_corpus but never broken into exemplars, so the richest style
channel never grew (8137/8126/8174 had 0 exemplars). The writer kept retrieving
only March–April seed paragraphs no matter how many finals were signed.
Extract the per-decision exemplar logic (section→paragraph→Voyage-embed→replace)
into a shared service `legal_mcp.services.style_exemplars.extract_and_store` —
the SINGLE implementation now used by BOTH the one-time backfill and the live
enrollment path (G2; no parallel extractor). `_enroll_final_in_library` calls it
on every final upload (source='internal_committee', the same source the writer's
search_style_exemplars reads). Voyage embeds over REST → container-safe;
best-effort, surfaced in the upload response, never fails the upload.
Effect: every signed final now grows the exemplar corpus, so the writer's
block-level style retrieval improves with each decision — the core "learn from
every decision" fix for channel B. Path A (style_distance_history) will track
whether the larger exemplar pool reduces style-distance over time.
Invariants: G2 (one extraction path shared by backfill + enroll), INV-LRN5
(style/structure prose only — substance routes elsewhere), INV-LRN4 (the
draft↔final loop now feeds the exemplar channel, not just the lesson channel).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The learning loop measured anti-patterns but never corrected them. style_distance
detects markdown headers / bullet lists / mid-paragraph mini-lists (from the
canonical lessons.ANTI_PATTERNS), but the writer only received anti-pattern
guidance if a chair `anti_patterns` override existed in appeal_type_rules — and
none does. So _build_style_context's `if ov:` branch injected nothing, the writer
was never told to avoid them, and drafts kept emitting them (8137: 28 hits, the
worst, newest — anti-patterns were trending UP, not down).
Anti-patterns are structural invariants of Dafna's voice (continuous legal
narrative — no markdown, no bullets), not overridable preferences. So inject the
canonical ANTI_PATTERNS notes ALWAYS, from the same list style_distance measures
against (single source of truth), with any chair additions layered on top. This
closes the measure-but-don't-correct gap: the next draft should show the markdown/
bullet anti-patterns drop, and Path A (style_distance_history) will confirm it.
Invariants: G1 (correct at source — the writer, not a post-hoc stripper), G2
(one canonical anti-pattern list shared by detection and instruction — no parallel
list), INV-LRN4 (closes the feedback half of the draft↔final loop).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
תיקוני-מבנה ללולאת-הלמידה (TaskMaster #157):
1. כרטיס ה-curator (/training טאב "אוצֵר") ספר decision_lessons WHERE source='curator'
— ערך שאף מסלול-קוד לא כתב → תמיד 0. עוצב-מחדש (דרך שער-העיצוב Claude Design, אושר)
להציג ביושר את שלושת ערוצי-ההזנה לכותב: דיסטילציה→appeal_type_rules (180, זורם),
פאנל→decision_lessons (81, ממתין), אוצֵר→source='curator'. get_curator_stats שוכתב.
2. drift ספ↔סוכן: 07-learning.md §1.1 + INV-LRN3 קבעו שהאוצֵר רושם ממצאים כ-decision_lesson
source='curator', אך הסוכן כתב comments בלבד — הממצאים האיכותיים אבדו. נוסף כלי-MCP
record_curator_findings + §A.5b ב-hermes-curator.md (read-only נשמר; הצעה מגודרת-שער).
3. get_recent_decision_lessons(limit=15) חתך בשקט — נוסף WARN על מה שנחתך (חוקה §6);
הפתרון האמיתי = סינתזת-לקחים (TaskMaster #158).
Invariants: מקיים INV-LRN1/G10 (שער-יו"ר), INV-LRN3 (לכידה מובנית), INV-IA2/IA5 (מקור-אמת
יחיד), G12 (leak-guard עובר), G2 (אין מסלול מקביל — אותם מאגרים). פער-מימוש פתוח מתועד:
§A לא רץ אוטומטית על mark-final (pipeline-wake exits) → נדחה לתכנון נפרד.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
A clean held-out test of voice learning was no longer runnable: every
final-uploaded case already has its lessons folded, and lessons are stored
universal/untagged so leave-one-out is impossible. Path A makes the test
prospective instead — capture the generalization datapoint at the one moment
it's clean.
On final upload, after the draft↔final pair is created but BEFORE this
case's lessons are folded (folding is a separate manual /training step), we
snapshot style_distance (anti_pattern_total, golden-ratio max-deviation,
change_percent) alongside the current voice-lesson pool size. Because the
draft was written with only the PRIOR pool, each row is a clean "with N
accumulated lessons, our draft on this unseen case scored X" datapoint. As
the pool grows over cases, a downward trend = learning generalizes.
- db: SCHEMA_V45 style_distance_history (append-only) + helpers
voice_lesson_pool_sizes / record_style_distance_snapshot /
get_style_distance_history.
- app: best-effort capture in api_upload_final_decision (never fails the
upload); GET /api/learning/style-distance-history for the trend.
Reuses the existing style_distance service + appeal_type_rules pool — no
parallel metric path. The 8 existing cases are already folded, so the table
starts empty and fills from the next final (their clean window is past).
Invariants: G2 (reuse style_distance/appeal_type_rules — one path),
INV-LRN4 (measure the draft↔final gap; this is its trend surface). LLM-free
(style_distance is deterministic) so it runs in the container.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
When a chair's signed final decision is uploaded it is enrolled into the
precedent library, but _enroll_final_in_library only set deterministic
fields (citation, proceeding_type, date) and copied subject_tags from the
case's subject_categories — which is usually empty. The Gemini metadata
pass (subject_tags / summary / headnote / key_quote) was never triggered,
so the row sat at metadata_extraction_status='pending' with no subject
tags until a drain happened to pick it up. In practice a freshly-enrolled
final showed an empty "תגיות נושא" in the precedent edit UI.
Fix: after enrollment + citation, call the existing reextract_metadata
path inline (G2 — one path, full status lifecycle). It runs on Gemini
Flash over REST (GOOGLE_GEMINI_API_KEY, already in Coolify), so it is
container-safe — unlike the halacha path (claude CLI, host-only).
apply_to_record fills only empty fields, so the deterministic seeds and
any chair-curated subject_categories are preserved. Result surfaced in the
upload response (out["metadata"]); failures logged, upload still succeeds.
Also corrects two now-stale comments: the "Gemini returns no_metadata for
internal decisions" note in the enroll loop, and the "MCP-tool-only path"
docstring on reextract_metadata (true for halacha, not for Gemini REST).
Invariants: G1 (fill at source on enroll, not a read-time workaround),
G2 (reuse reextract_metadata — no parallel extraction path). LLM call is
Gemini REST, not claude_session, so the container LLM-call constraint holds.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Inline # comments in gitignore are not supported — they were silently
breaking three patterns (data/checkpoints/, data/adapter-migration-state.json,
.claude/agents/.generated/). Moved comments to their own lines and added
missing entries for runtime dirs (data/audit/, data/logs/, etc.) and
temp files (.interaction_tmp.json, .design-build/, .taskmaster bak files).
Also tracks previously untracked legitimate files: scripts, tests, docs,
skills references, .env.example, taskmaster templates.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Switches the scanned-PDF fallback from Google Cloud Vision to
Mistral OCR (mistral-ocr-latest) for better Hebrew accuracy and
robustness against broken embedded OCR layers (e.g. case 1044-03-26
which returned English garbage through Vision).
Routing strategy (document-level, not per-page):
- PyMuPDF extracts all pages; pages that pass _text_quality_ok()
use PyMuPDF output directly (free, ~50ms).
- If ANY page fails quality → Mistral OCR called once for the whole
PDF, returning per-page Markdown for all pages (consistent format,
no plain-text/Markdown mix within a document).
Markdown output preserved: Mistral returns ## headers and |tables|;
chunker updated to recognise ATX Markdown headers (##/###) as section
boundaries in _split_into_sections().
Config: GOOGLE_CLOUD_VISION_API_KEY → MISTRAL_API_KEY; allowlist
updated vision.googleapis.com → api.mistral.ai.
MISTRAL_API_KEY added to Coolify container env.
Invariants: G1 (single OCR fallback path, not parallel), G2 (no
duplicate extractor route), INV-AH (Mistral handles gershayim
natively; quote-fix only on PyMuPDF path).
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
aggregate_claims_to_arguments crashed with "current transaction is aborted,
commands ignored until end of transaction block" on large cases (confirmed on
1027-04-26, 195 claims; reported via CMP-186).
Root cause: the proposition INSERT (legal_argument_propositions) was wrapped in
a broad except Exception with no savepoint. When the LLM echoes a
syntactically-valid-but-nonexistent claim_id, the FK violation
(legal_argument_propositions_claim_id_fkey) puts the asyncpg transaction into
aborted state. The except caught only that INSERT's error but never issued
ROLLBACK TO SAVEPOINT, so the next statement (the following argument's INSERT
into legal_arguments) raised InFailedSQLTransactionError uncaught and crashed
the whole call. With many claims the bad-UUID probability is high -> consistent
failure.
Fix:
- Validate each claim_id against the known set of claim ids fetched for the
case before INSERT, so a hallucinated id never reaches the DB (G1: fix at
source). Malformed UUIDs are already dropped in _normalize_argument; this
catches the valid-but-nonexistent ones.
- Wrap the INSERT in a per-row savepoint (async with conn.transaction()) as
defense-in-depth, so any unexpected constraint failure rolls back to the
savepoint instead of poisoning the surrounding transaction.
Invariants: G1 (normalize at source, not symptom-catch on read). No parallel
path (G2). No silent swallow — skips are logged.
TaskMaster: legal-ai #156
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Backend half of the citation-verification tab (frontend follows in a separate PR):
- Schema V44: case_precedents gains argument_id (the legal argument it supports),
case_law_id (the corpus ruling — for cited_by + dedup), verified (the INV-AH gate
the writer respects) + verified_at. All nullable; chair_note already existed.
- db: create_case_precedent(+argument_id/case_law_id/verified),
set_case_precedent_verified(id, verified, chair_note), list_case_precedents(+cols).
- services/case_citation_verification.build_view(case): per legal_argument →
in-corpus suggestions (search_library per issue) each with the cited_by authority
breakdown (db.citation_authority, X11) + merged attach/verify state + per-issue
radar (case_digest_radar grouped by matched issue). Pure read/assembly; reuses the
one corpus search + one authority query + one radar (G2).
- endpoints: GET /api/cases/{n}/citation-verification (the view) and
POST .../verify (upsert verify/un-verify + chair note).
Validated on 1044-03-26: 14 arguments, all with support; 317/10 surfaces with
אומץ×11/אובחן×1; radar leads attributed per issue. Auto-approval untouched (chair
gate, INV-G10). Verify defaults to false — nothing is authoritative until the chair
marks it.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The radar's query was built from dozens of raw claims (procedural-heavy noise), so
matches were thematic but imprecise and gave no reason WHY a lead is relevant. Now:
- Prefer the analyst's distilled legal_arguments (argument_title + legal_topic — one
crisp CREAC issue per row) over raw claims.
- Search EACH issue separately and MERGE, so every lead is attributed to the case
issue(s) it answers (`matched_issues`) — the chair sees "this ruling is for your
'זכות עמידה' issue", not just a blended score.
- Fall back to the raw-claims blended query pre-aggregation; `source` reports the path.
- Shared `_radar_enrich` helper (gap status + action + matched_issues), bounded to 25
issues to cap the per-issue fan-out.
Validated: 8124-09-24 (32 args → per-issue) surfaces betterment rulings each tagged to
its issue (היעדר השבחה / זהות הנישום / סעיף 7(ב)); 1044-03-26 (0 args) falls back to
claims unchanged. No tool/endpoint signature change (new fields pass through the dict).
Invariants: G2 (reuses the one digest search + arg accessor), INV-DIG1 (radar only).
No schema change.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
A digest pointing at a ruling we don't hold yet ("unlinked") was captured globally
(missing_precedents inbox) but never surfaced IN THE CONTEXT of the case being
decided — so a relevant ruling known only via a digest could fall through the cracks
at the moment it matters. Adds the case-contextual radar:
- db.search_digests_semantic: new `linked_only` filter (False = unlinked-only target set).
- digest_library.case_digest_radar(case_number): builds the case topic from title +
appeal_subtype + the analyst's claims, embeds once, matches against UNLINKED digests,
and returns leads enriched with the underlying ruling's gap status + suggested action
(new_lead / gap_open / fetched / available_link).
- MCP tool `digest_radar` + endpoint GET /api/cases/{n}/digest-radar (for the agent and
the future case-page lead).
INV-DIG1 preserved: radar only — every lead points at the underlying RULING (fetch /
upload / link), never cites the digest. Read-only.
Validated on 8124-09-24 (היטל השבחה): 5 on-topic unlinked digests, scores 0.64–0.72.
The visible case-page panel is a separate UI change → goes through the design gate.
Invariants: G2 (reuses the one digest semantic-search + gap/citation resolvers),
INV-DIG1 (no digest citation), INV-DIG3 (gap surfacing). No schema change.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The internal citation graph fed only RANKING (raw in-degree), and the per-citation
TREATMENT was never classified — so a precedent distinguished N times got the same
authority boost as one followed N times (INV-COR2 violation), and the signal never
reached the agents' reasoning. Wires the full path:
Phase 1 — scripts/classify_citation_treatments.py: classify each linked edge's
treatment (followed/distinguished/…) from its match_context via
corroboration.classify_treatment (Opus 4.8 @ xhigh, local), filling
precedent_internal_citations.treatment. Idempotent.
Phase 2 — db.refresh_verified_layer: count only NON-negative treatments toward
verified/cite_count (INV-COR2/COR4). Unclassified counts as neutral-positive so
the signal degrades gracefully before classification runs.
Phase 3 — db.citation_authority(ids): per-precedent {total, positive, negative,
unclassified, by_treatment}. Surfaced as `cited_by` in search_precedent_library
hits and precedent_library_get, and `treatment` per incoming citation.
Phase 4 — legal-researcher/analyst/writer prompts: weigh & ARGUE authority
("הלכה שאומצה ב-N החלטות ועדת-ערר"), flag distinguished/overruled, never invent
the count (INV-AH; writer is read-only of the analyst).
Auto-approval stays kill-switched off (chair gate preserved, INV-G10). No schema
change (treatment column already existed). Operational: run the classifier +
refresh_verified_layer over the 379 edges, then sync agents across companies.
Invariants: G2 (one classifier + one authority query, reused), INV-COR2/COR3/COR4
(negative never corroborates; point-specific; ≥N), INV-G10 (no auto-approval),
INV-AH (no invented numbers).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Root cause: the upload form "מראה המקום" field is stored only as case_number;
citation_formatted was left empty and filled ONLY by the async metadata
extraction, which runs in the local drainer (not the container) because it was
lumped with halacha extraction. Result: the chair saw an empty מראה-מקום during
the window and re-typed it by hand.
Fix (hybrid — Option C):
- Seed: _create_external_record / create_external_case_law store the chair's typed
citation as citation_formatted at INSERT, so the field is never blank. A
cited_only→external promotion preserves an existing non-empty value (prior edit).
- Enrich: ingest_precedent runs metadata extraction INLINE in-container
(gemini_session is direct REST, no local CLI) with force_citation=True, upgrading
the seed to the canonical derived citation (parties + reporter + date) before the
chair opens the page. Best-effort: on no key / API failure the seed remains and
the queued metadata drain stays as the fallback. Halacha extraction is untouched
(stays local — claude_session).
- apply_to_record / extract_and_apply gain force_citation: re-assemble even when
citation_formatted is non-empty, writing only when assembly SUCCEEDS (seed
preserved on abstention). The drainer keeps the default (False) so a chair's
manual edit in /precedents/[id] is never clobbered.
Ops: GOOGLE_GEMINI_API_KEY added to the legal-ai Coolify app (runtime) — the SoT
value from Infisical nautilus:/external-apis/gemini, validated against the API.
Invariants: G2 (one citation-resolution + one metadata-extraction path, reused —
no parallel logic), INV-ID2/X1§3 (citation_formatted stays a derived field; the
seed is provisional and upgraded), INV-G10 (chair edits preserved). No schema change.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
