vignette/docs/phase3/dataset-export-manifest.md
Yun Chan 085460b5e0 대시보드 폴드아웃/드릴다운 정리 + 페르소나 역린·misconduct 반응 + 게이트웨이 격리·RAG 비차단 수정
SSOT 대시보드:
- 한신대 기술분석 PDF(19쪽) 정합성 분석 + 이번 세션 발견 섹션 추가
- 섹션 폴드아웃(접기)·상단 목차(드릴다운)·모두 펼치기/접기 — 내용 보존, 레이아웃만 정리

페르소나 반응 강화('저항·반응 조절' 핵심 차별):
- PersonaCard.triggers(역린) 필드 + CCD 핵심상처 파생 역린 블록
- L0에 무례·모욕·조롱 시 현실적 동맹 균열 반응 지침

버그·성능 수정(라이브/E2E로 포착):
- 게이트웨이 페르소나 격리: --append-system-prompt를 --system-prompt(교체)로 + --exclude-dynamic-system-prompt-sections (내담자 캐릭터 붕괴·개발맥락 누출 차단)
- RAG: 임베더 동기 로드(약 7-13초)를 _warm_rag_caches 백그라운드 warm으로(세션 생성 블로킹 회귀 수정)
- voice TTS RMS 데드힌트 제거, init_state OpennessParams 파라미터객체화
- 한국어 PII(날짜·금액·주소) 마스킹 보강
- 레이아웃 시각 게이트: 폼 컨트롤 값 스크롤 오탐 제외(7/7)

검증: 백엔드 84/84, E2E 42(데스크톱 27·모바일 11·아바타 4), 시각 게이트 7/7
2026-06-27 02:30:46 +09:00

4.8 KiB

Phase 3 Anonymized Dataset Export Manifest

Purpose

Define the manifest and JSONL expectations for the first recursive-learning dataset artifact. The export is allowed only after consent scope, withdrawal state, PII masking, and reviewer agreement are documented.

Primary evidence file: 03-export/export_manifest.json

Export Gate

The export may be marked approved_for_recursive_learning_seed only when:

  • Source participants are active or otherwise allowed by consent scope at export time.
  • Withdrawn participants are excluded according to the approved withdrawal policy.
  • Direct identifiers and raw identity mapping are absent from exported files.
  • Free text is masked or transformed according to the manifest.
  • PII scan is pass, or every finding has a reviewer disposition.
  • IAA gates pass for gold exports: kappa >=0.60 and ICC >=0.75.
  • Legal/data-steward review is complete.

If any gate is missing, set export_status to technical_dry_run or blocked, not approved_for_recursive_learning_seed.

Manifest Shape

{
  "export_id": "phase3-rl-seed-YYYYMMDD-001",
  "dataset_name": "vignette_phase3_recursive_learning_seed",
  "export_status": "technical_dry_run",
  "created_at": "2026-06-26T00:00:00Z",
  "purpose": "recursive-learning seed dataset for education simulator improvement",
  "source_window": {
    "started_at": "2026-00-00T00:00:00Z",
    "ended_at": "2026-00-00T00:00:00Z"
  },
  "source_tables": [
    "app.sessions",
    "app.turns",
    "app.feedback_scores",
    "ds.annotation",
    "audit.persona_drift_log"
  ],
  "selection_criteria": {
    "cohort_id": "phase3",
    "min_completed_sessions": 2,
    "include_withdrawn": false,
    "excluded_safety_scope": ["self_harm_scenario_primary"]
  },
  "consent_scope": {
    "consent_version": "",
    "allowed_uses": ["education_quality_review", "recursive_learning_seed"],
    "withdrawal_cutoff_applied_at": "",
    "participants_included": 0,
    "participants_excluded": 0
  },
  "anonymization": {
    "participant_key": "pseudonymous export key; no identity map included",
    "text_transform": "masked_text_only",
    "direct_identifier_policy": "blocked",
    "salt_or_identity_map_location": "not in export"
  },
  "pii_scan": {
    "tool": "",
    "version": "",
    "ran_at": "",
    "status": "pending",
    "findings": []
  },
  "agreement": {
    "kappa": null,
    "icc": null,
    "gold_status": "not_gold"
  },
  "files": [
    {
      "path": "03-export/anonymized_dataset.jsonl",
      "rows": 0,
      "sha256": "",
      "schema": "phase3_dataset_item_v1"
    }
  ],
  "approvals": {
    "data_steward": "",
    "legal_or_privacy_reviewer": "",
    "technical_operator": "",
    "approved_at": ""
  },
  "known_limitations": []
}

JSONL Record Shape

Each JSONL line must be one dataset item:

{
  "schema": "phase3_dataset_item_v1",
  "item_id": "export-local-id",
  "participant_key": "PX-0001",
  "session_key": "SX-0001",
  "turn_key": "TX-0001",
  "persona_id": "P1",
  "stage": "rapport",
  "speaker": "counselor",
  "text_masked": "masked or transformed utterance text",
  "techniques": [],
  "client_states": [],
  "feedback_scores": [],
  "supervisor_comments": [],
  "source_refs": {
    "session_started_at": "2026-00-00T00:00:00Z",
    "export_manifest_id": "phase3-rl-seed-YYYYMMDD-001"
  },
  "privacy": {
    "direct_identifiers_removed": true,
    "pii_scan_status": "pass",
    "consent_scope": "recursive_learning_seed"
  }
}

Blocked fields:

  • name
  • email
  • phone
  • student_id
  • national_id
  • address
  • date_of_birth
  • raw_audio_path
  • raw_voice
  • raw_source_case
  • identity_map
  • API keys, tokens, cookies, or credentials

Operator Commands

Generate the export only through an approved read-only reporting job. Until that job is implemented, use the manifest as a contract and mark exports technical_dry_run.

Validate artifact presence and headers:

python scripts\check-phase3-artifacts.py --check --evidence-root $root

Write a validation report:

python scripts\check-phase3-artifacts.py --check --evidence-root $root --output "$root\phase3_artifact_check.json"

Compute a file hash for the manifest:

Get-FileHash "$root\03-export\anonymized_dataset.jsonl" -Algorithm SHA256

Export Review Checklist

  • Manifest export_status matches the actual approval state.
  • Files listed in files[] exist and have row counts and SHA256 hashes.
  • participants_included plus participants_excluded matches the consent roster decision log.
  • include_withdrawn is false unless legal review explicitly approved otherwise.
  • Text fields use masked/transformed text only.
  • PII scan report is attached or referenced.
  • Agreement metrics match 02-measures/kpi_report.json.
  • Export limitations state whether the file is gold data or a dry-run seed.