R5 = approval-gated read/report round (docs/41 §6.3 Round 5). Codex gpt-5.6-sol wrote tools/r5_dry_run_report.py + docs/42_r5_dry_run_impact_report.md; I (director) measured ground truth, directed, and verified. Branch feat/ontology-brain-split, HEAD f705598 unchanged, NOT committed. All read-only: SQLite mode=ro/query_only, Neo4j MATCH/RETURN only, extracted text never persisted, Notion block-reads only. Zero production writes confirmed (sqlite-wal 0 bytes, no commit).

Key enabler: stored Drive file_ids are stale (404), so live re-extraction used the local cache /home/insu/rosaic/GoogleDrive.bak (mirrors drive tree). 102/171 stubs matched by exact basename; ran the real R2 extractors (rka.ingest.binary_text) read-only.

MEASURED (script cross-checked director ground truth, 0 discrepancies): 1. Binary recovery: 171 stubs, 102 matched. 40 recover real text (845,255 chars) — hwp 27 (median 5,260), hwpx 6 (median 32,180), pdf 1 (251,344), docx 1 (109,470), pptx 3, txt 2. Honest non-recovery: 25 pdf ocr_required, 19 no_embedded_text, 17 pptx skipped_resource_limit (>50MB cap), 1 hwp low_text_yield. 69 unmatched projected (59 resource-capped pptx, 9 attemptable, 1 unsupported). 2. Ghost disposition: structured_gdrive=235 (all folder-path echoes, 0 human-approved) -> 47 review-only revivable (same-source real mention), 137 archive (no mention), 51 review-only content-unavailable. 16 stub-fed llm_new -> archive/quarantine all. 3. Defect ③ ceiling: currently 40/2640 (1.5%) Problem/Decision linked to Project/Person. Generous co-occurrence ceiling (content cue + Project/Person surface + relation cue in one window) = 13 qualifying docs -> at most 53/2640 (2.0%). Real C.4-validated edges a fraction of that. "Not magic" quantified. 4. Workstream embeddings: 12 nodes (3 missing + 9 stale), all others skipped (6016). Missing: [Operational Nervous System], [UMI] Prior research survey, [VFLA] Improve Model. Clean derived-index rebuild. 5. Person dedup: 21 nodes, 5 groups x 3 = 15 (1 notion canonical + 2 rka ghosts each) -> 11 nodes. Ghost blast radius 43 edges total (per group one degree-0 ghost + one connected: 임동성 17, 이호준 10, 문승연 9, 권우현 5, 김인수 2). 6 singletons untouched. 6. Mirror digest (R3): 324 notion_mirror, currently 310 empty (max 225 chars). 236 linkable entity pages in R3 scope (88 temporal facts out of path). Live Notion sample 48/48 succeeded (overturned my empty-DB-rows assumption): 25/48 have >=600-char narrative -> projected ~123 of 236 eligible pages materially grow. 2400 is a cap not a promise.

R6-R8 plan in docs/42 §8: 7 ordered approval-gated steps. "Backfill is not cleanup" — source re-extraction (R6/A2), suggestion reclassification (R7), Neo4j graph reconciliation (R7) are THREE separate reversible state transitions each needing own snapshot/rollback. Blocked owner decisions: OCR scope (25 pdf + 19 no-text), on-demand queue for 76 over-cap stubs. R6 backfill must re-list drive + map old->new by path/name/size/mime/ancestry (GoogleDrive.bak is measurement proxy only, not prod source).

Verify: ruff clean; pytest INDEPENDENTLY re-run by me = 759 passed, 3 failed (only pre-existing test_source_config_seed, unrelated to R2 max_binary_bytes seed). Lesson: codex exec long-runs exceed 10-min foreground Bash cap -> run in background with logfile + poll; a SIGTERM'd foreground codex leaves the detached node process alive (kept running read-only 12+ min) -> kill orphan PIDs before trusting the final artifact.

Sagwan Revalidation 2026-07-16T01:20:59Z#

• verdict: refresh
• note: HEAD/미커밋/로컬 캐시 등 시점 의존 주장이 많아 재측정 필요

Sagwan Revalidation 2026-07-18T02:35:25Z#

• verdict: ok
• note: 측정 보고로 시점 의존 외부 주장 적고, 최근 검증 후 변경 징후 없음.

Sagwan Revalidation 2026-07-20T03:25:25Z#

• verdict: ok
• note: 최근 검증 후 변동 근거 없고 측정·읽기전용 주장도 재사용 가능