G0~G8 성과·동맹 측정 OS 작업 일괄 고정

8월 7일까지 워킹트리에만 남아 있던 미커밋 작업을 커밋한다. 여러 사본
폴더(worktree·clone)에 흩어져 있던 중간 스냅샷을 정리하기 전에 원본을
git 이력으로 고정하는 것이 목적이다.

- contracts/routes/services: measurement, outcome_trajectory, rupture_repair,
  deliberate_practice, calibration_transfer, supervision_research,
  multimodal_alliance, continuous_improvement 계열 신규 모듈과 테스트
- infra/db/init: 07~16 마이그레이션(측정 기반~calibration transfer 실행)
- apps/web: 세션 리뷰 카드·관리 화면·E2E 스펙 추가
- docs/ops: G0~G8 라이브 통합·배포·롤백 증거 문서와 evidence JSON/PNG
- scripts: smoke·ledger·릴리스 에이전트·NAS 프리뷰 운영 스크립트

engine.public 로그 .bak과 apps/web/test-results 산출물은 커밋에서 제외했다.
This commit is contained in:
Yun Chan 2026-08-08 01:30:53 +09:00
parent 93dd8f82d7
commit 16e791e044
390 changed files with 243188 additions and 499 deletions

View file

@ -163,6 +163,15 @@ _PII_PATTERNS: list[tuple[str, re.Pattern[str]]] = [
("ADDR", re.compile(r"[가-힣]{2,}(?:시|도)\s?[가-힣]{1,4}(?:시|군|구)\s?[가-힣0-9]{1,}(?:동|읍|면|로|길)")),
]
# Source/input text keeps the broad contextless-name heuristic above for recall.
# Generated synthetic content has no user-originated text after source validation;
# applying that heuristic there misclassifies ordinary words such as "서운함을".
# Keep every high-confidence label/context/honorific pattern and all non-name PII.
_SYNTHETIC_GENERATED_PII_PATTERNS = (
*_PII_PATTERNS[:5],
*_PII_PATTERNS[6:],
)
# Presidio 지연 로드 캐시 (-1=미시도, None=미설치, 객체=설치됨)
_PRESIDIO_ANALYZER: object = -1
_PRESIDIO_ANONYMIZER: object = -1
@ -216,7 +225,10 @@ class MaskResult:
used_ko_recognizer: bool = False
def _mask_regex_pii(text: str) -> tuple[str, list[str]]:
def _mask_regex_pii(
text: str,
patterns: Iterable[tuple[str, re.Pattern[str]]] = _PII_PATTERNS,
) -> tuple[str, list[str]]:
masked = text
found: list[str] = []
@ -235,7 +247,7 @@ def _mask_regex_pii(text: str) -> tuple[str, list[str]]:
return _replace
for label, pat in _PII_PATTERNS:
for label, pat in patterns:
masked = pat.sub(replace_match(label), masked)
return masked, sorted(set(found))
@ -307,6 +319,50 @@ def mask_pii(text: str) -> MaskResult:
)
def mask_synthetic_generated_pii(text: str) -> MaskResult:
"""PII gate for model-generated text from validated synthetic sources.
Explicit name labels, self-introductions, relationship/name contexts,
honorifics, optional recognizers, Presidio and every non-name PII pattern stay
enabled. Only the ambiguous contextless Korean surname heuristic is omitted.
"""
if not text:
return MaskResult(text_masked=text, entities=[], used_presidio=False)
analyzer, anonymizer = _try_load_presidio()
if analyzer is not None and anonymizer is not None:
try:
results = analyzer.analyze(text=text, language="en")
ents = sorted({r.entity_type for r in results})
anonymized = anonymizer.anonymize(text=text, analyzer_results=results)
ko_masked, ko_ents, used_ko = _mask_ko_recognizer_pii(anonymized.text)
masked, regex_ents = _mask_regex_pii(
ko_masked,
_SYNTHETIC_GENERATED_PII_PATTERNS,
)
return MaskResult(
text_masked=masked,
entities=sorted(set(ents + ko_ents + regex_ents)),
used_presidio=True,
used_ko_recognizer=used_ko,
)
except Exception:
pass
ko_masked, ko_ents, used_ko = _mask_ko_recognizer_pii(text)
masked, found = _mask_regex_pii(
ko_masked,
_SYNTHETIC_GENERATED_PII_PATTERNS,
)
return MaskResult(
text_masked=masked,
entities=sorted(set(ko_ents + found)),
used_presidio=False,
used_ko_recognizer=used_ko,
)
# ════════════════════════════════════════════════════════════════════════════
# 2. 위기 분류 (입력 — 실제위기 vs 페르소나 연기 구분, R8)
# ════════════════════════════════════════════════════════════════════════════