두 갈래의 작업이 한 커밋에 들어 있다. llm_assistance.py 에서 T0-4 가 앞선
작업의 프롬프트 수정을 대체했기 때문에(프롬프트 리터럴이 llm_prompts.py 로
이동), 분리하려면 실제로 존재한 적 없는 중간 상태를 만들어야 한다.
## (1) 앞서 진행 중이던 review-intelligence 작업
- decision_stats(): 위험밴드 x 실제판정 교차표 + 캘리브레이션 지표
- cleared_reference 지식엔트리 타입: 이전에 클리어된 이미지 매칭은 +80 이
아니라 0점 긍정 신호
- LLM 요약 본문 영속화(_evidence_payload 에 summary/sourceUrls) — 그전까지
요약이 저장되지 않아 운영자에게 한 번도 보인 적이 없었다
- 운영자 GUI 반영
## (2) 실사용 투입 전 T0 — 첫 판정 이후에는 소급 불가한 것들
- T0-1 initialRiskScore/initialRiskBand/scorerVersion 각인.
_rescore_submission 이 12곳에서 riskScore 를 덮어쓰므로 저장된 점수는 늘
왕복 점수였고, 그것을 재채점과 비교하는 테스트는 항상 통과하는 빈 테스트다.
레거시 행은 scorerVersion="legacy" 로 표시만 하고 점수를 날조하지 않는다.
- T0-2 얼굴검출 Haar -> YuNet, 인물 등급(identifiable/background) 도입.
기존의 평평한 +35 는 사람이 찍힌 모든 사진을 medium 문턱 위로 올려 밴드
분포를 붕괴시켰다(실측 high 4 / medium 2 / low 0). 등급은 payload 라운드
트립으로 보존한다 — 잃으면 재채점마다 조용히 +35 로 되돌아간다.
가중치는 models/(gitignore) 에 두고 offline-install.md 에 sha256 반입 절차.
가중치가 없으면 Haar 로 강등하고 1회만 경고한다.
- T0-3 판정 시점 피처 스냅샷. 큐가 riskScore 내림차순이고 화면이 점수를 먼저
보여준 뒤 판정을 받으므로, 스냅샷 없이 모은 라벨은 "이 근거가 반려를
예측한다"가 아니라 "운영자가 내가 보여준 점수에 동의했다"를 기록한다.
- T0-4 모델/프롬프트 버전 + latency_ms 스탬프. latency_ms 는 _evidence_id
해시에서 제외한다 — 포함하면 같은 요약이 재생성될 때마다 id 가 바뀐다.
## (3) 임베딩 저장 금지 정책 개정
금지의 취지는 생체 식별이지 이미지 유사도가 아니므로 범위를 좁힌다.
장면 임베딩(전체 이미지 의미 벡터)을 이미지 유사도 판정에 한해 허용하되
얼굴 마스킹 / 인물 지배 프레임 제외 / 목적 제한 / 별도 저장 클래스 /
연쇄 삭제 5조건을 assert_scene_embedding_allowed 로 강제한다. 마스킹 증명이
없으면 거부한다 — 누락은 마스킹을 건너뛴 것과 구별할 수 없다.
얼굴 임베딩 금지는 landmarks/face_descriptor/face_encoding 추가로 강화.
pytest 502 passed (착수 전 463).
- Keyboard review mode: J/K to move, A/H to decide and jump to the next
unreviewed case, R to focus the memo, N for next unreviewed, ? for help, with
a live position/backlog indicator; disabled while typing.
- Batch decisions: POST /api/submissions/decision-batch applies one decision to
many selected cases through per-case transactions/audit/watchlist, reports
per-case success and failure, caps at 200, requires a memo for reject and
confirms when high-risk cases are included. Lifts the v1 "no bulk
approve/reject" spec restriction under those guardrails (spec revised, static
layout test updated to assert the guardrails instead of the absence).
- Decision-memo presets: GET/POST /api/decision-memo-presets (upsert by label),
one-click chips in the decision pane, "save current memo" action.
- Submission duplicate / prior-case index: seed backfills fingerprints onto the
submission record (idempotent), exact + pHash>=0.97 clustering, queue badge
and group select, prior-decision chips and a conflicting-decision flag.
- Decision export: GET /api/decisions/export with separate internal and
applicant field whitelists (applicant view never carries score/evidence),
CSV with a BOM.
- Online backup/restore: POST /api/admin/backup snapshots the WAL database via
the sqlite backup API plus the media dirs into a timestamped zip with
retention rotation; `--restore` is an offline CLI that keeps the pre-restore
database. GET /api/admin/backups lists snapshots.
Also folds in the co-located HTTP hardening (non-object/non-UTF8 JSON body ->
400, non-ASCII Authorization/cookie compared as bytes -> 401, media OSError ->
404) and the line-buffered startup logging that share these files.
Verified against a live server (auth gate, seeding, duplicate index, batch
decision guardrails, presets, export, backup + restore round-trip). pytest
tests/rights_filter tests/operator_gui: 457 passed.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01E6GD9gxUtfUiVB3JHejGpU
- Image viewer controls (fit / 1:1 / rotate) now have handlers and reset per
case (were dead buttons).
- 'auto-entry deactivate' is scoped to the selected case's sourceSubmissionId
(was deactivating an arbitrary global entry).
- dismissCollectionCandidate persists via a real endpoint (was a local splice
that resurrected on the next refresh).
- setDecision has an in-flight guard; rerunSelectedEnrichment /
applyCoverageFilter get null guards and a best-effort refresh on partial
failure; deactivateCorrectionEntry skips already-inactive entries, tolerates
per-entry failures, and always refreshes (guarded).
- executeSuggestedQueries / runCandidateCollection commit only after success
so a failed run doesn't wipe the previous view; applyBootstrap clears a stale
error banner; renderCorrections guards derivedEntries.
- riskClass allowlists the band and case/similar/face-crop URLs route through
safeUrl (XSS defence-in-depth).
Backend (data-integrity / audit-trail):
- mark_evidence_status: make the read-modify-write + audit atomic under
_write_lock + _transaction (was 4 unlocked connections, last-writer-wins,
audit could diverge from DB state).
- promote_collection_candidates: acquire _write_lock, use a deterministic
_stable_id, wrap writes in _transaction, and guard already-promoted batches
so a retry/double-click cannot mint duplicate knowledge entries.
- risk_scoring: bucket web-detection entities by name so two distinct entities
in one image both count (was collapsed to one 'entity' bucket, under-scoring
high->medium). Identical entities still dedup.
Frontend (silent loss of operator actions):
- runManualSearch: in-flight guard; rerunHistoricalQuery voids the call,
preventing duplicate/racing search POSTs.
- deactivateCorrectionEntry: persist via the real knowledge-deactivate
endpoint instead of a local-only mutation + fake audit event.
- provider-retry button: report honestly (no backend endpoint) instead of
writing a fake 'Provider called' audit event.
Tests: +4 regression tests (mark_evidence_status rollback, batch-promote
idempotency, two distinct entities -> high, identical entity dedup).
Remaining Phase 1 item #6 (_auto_naver/_auto_google refactor) deferred.
Backend suite: 361 passed. app.js: node --check OK, 41 static tests pass.
Address commit security review: the same-origin branch of safeUrl accepted
//host and /\host, which browsers normalize to an external host (open
redirect). Allow only true same-origin paths.
Add safeUrl() to gate external search-result URLs into href/src (blocks
javascript:/data:), parse the response body before the ok check in apiJson
so non-JSON error bodies surface the real status, and hide broken evidence
preview images via onerror.
- Bundle Pretendard Variable woff2 locally (air-gapped safe, no CDN)
and switch UI/stamp font stacks to it; preload in index.html
- Replace the forensic-dossier paper theme with a flat neutral cool
palette: single teal accent, white cards, no noise texture, and
zero linear/radial gradients (per design contract)
- Restore the product-purpose top-bar block and its CSS, drop the
unused global search form, and strip the stray UTF-8 BOM
- Re-skin queue hover/selection, eyebrows, nav rail, chips, and
empty states to the neutral palette; tabular numerals for numbers
- Regenerate ui-overhaul final audit artifacts: zero horizontal
overflow across 8 views at 1440x900 and 390x844, Pretendard active
Design spec: docs/superpowers/specs/2026-06-11-operator-console-clean-review-ui-design.md
Plan: docs/plans/2026-06-11-001-feat-operator-console-clean-review-ui-plan.md
Tests: 358 passed (full suite incl. browser smoke)
Image rights / copyright detection system: SQLite store, HTTP app,
search integrations (Naver, Google Custom Search, Google Cloud Vision
web detection), image analysis (fingerprints, face/person detection,
evidence enrichment, risk scoring), an admin/review layer, governance
and retention policies, batch jobs, and a browser-based operator GUI.
This baseline incorporates a full code-review remediation pass
(46 fixes; 358 tests passing). Highlights:
CRITICAL
- Prevent evidence cascade-delete during the schema-constraint
migration by disabling FK enforcement around the table rebuild.
Security
- Sandbox served media (neutralize stored XSS from uploaded/collected
SVGs) via CSP + nosniff on the untrusted media routes.
- Strip embedded EXIF/GPS from external image derivatives before they
are sent to third-party APIs.
- Return a clean 404 (not an uncaught StopIteration) for PATCH on an
unknown provider.
Correctness
- LLM-summary failures no longer add +30 to the risk score.
- Decode only explicit JS escapes so Korean image URLs are not mangled.
- Consume search quota only after a successful request.
- Naver/Google adapters map responses inside the failure boundary, so a
malformed response degrades to evidence instead of crashing enrichment.
- Domain-aware provider attribution; face-box IoU de-duplication; count
searches (not result items); per-box crop isolation; clamp evidence
confidence and Google CSE num; real submittedEpoch; and more.
Robustness
- Offline LLM connect fast-fails (short connect timeout) so seed/reload
requests are not stalled; full read timeout preserved for generation.
- Malformed numeric env vars fall back to defaults instead of crashing
startup.
Performance
- Per-submission evidence reads (no full-table scan per rescore),
audit-log LIMIT, lazy active-store lookup, hoisted timestamps.
Tests
- ~24 regression tests added pinning the above fixes.
Runtime data (data/, outputs/, *.sqlite3, *.log), secrets (.env), and
node_modules are gitignored.
