[CLAUDE] Docs: Harness-17 deep-reverify (WF3) — fault-injection teeth-proof + cross-doc + AI_INFRA update

- WF3 wf_6f44eddf-680 (anh-directed 3rd pass): Lane A fault-injection PASS (13 inject prove teeth: A1/A2/C-spectrum/B1/B2/A3 all catch faults, no-false-flag, off-by-one correct, 8/8 md5 match restore) + Lane C integration PASS (0 application-defect, anti-Goodhart flip proofs, numbering intact, mfe-eval A2 not broken)
- Lane B empty-return (#53) → em-main self-gate recover (N=3/HCV/run-id/hash consistent, detector 27)
- engine §I banner: inline WF2 wf_8e73c1d6-ee6 + WF3 wf_6f44eddf-680 (was WF2-pointer-only)
- adap-report: +Deep-reverify(WF3) section + 2 brittleness honest-noted (fail-safe over-flag, not fixed per Lane A) + dogfood-C3-round2
- email AI_INFRA update (hash 8c94fa348514): deep-reverify notify
- dogfood: WF3 reviewer self-write reverted (D9/AS-10 round-2); cicd-monitor verdict kept (legit normal-spawn) — Run #476 success, prod 200x3, Mig 59
- 0 .ps1 changed → CI-skip (docs-only); 0 production code

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
This commit is contained in:
pqhuy1987
2026-07-01 09:31:55 +07:00
parent 18a17f5498
commit 65b434a370
5 changed files with 60 additions and 7 deletions

View File

@ -4,7 +4,7 @@ from: ai_infra
project: se (SOLUTION_ERP)
date_applied: 2026-07-01 (S95)
status: EXECUTED + runtime-verified (spec-audit script) · VERIFIED-pending-restart (cadence command wire)
reviewer_gate: PASS (WF2 3-lane: code-gate + completeness/cross-harness + honesty, run-id wf_8e73c1d6-ee6)
reviewer_gate: PASS (WF2 wf_8e73c1d6-ee6 3-lane + WF3 wf_6f44eddf-680 deep-reverify fault-injection — both PASS 0-blocking)
hash_verify: MATCH both broadcasts (content_sha256 recompute = declared; PowerShell canonical)
---
@ -41,6 +41,20 @@ Adopt **Harness-17 = memory self-improvement loop (Eval → Refine → Audit)**
- **WF1 IMPLEMENT** `wf_f085d7c4-d78` (4-agent ground-truth: reuse-anchor + C1-audit-design + B1/B2/B3-draft + cross-harness-conflict) → em-main integrate single-writer. **Lesson dogfood:** WF1 Agent-A (investigator-codebase) tự-compact MEMORY.md giữa workflow (persona vượt return-only-prompt) = **AS-10/E-006 recurrence** → em-main git-diff bắt + **revert** (D9 single-writer). Bonus: `archive/2026-06.md` KHÔNG tồn-tại → compaction sẽ **mất 8.5KB** nếu giữ → **dogfood chính C3 floor đang adopt cứu data thật**.
- **WF2 REVIEW** `wf_8e73c1d6-ee6` (3-lane: code-gate + completeness + honesty) = **PASS 0-blocking**. **Anti-fiddle proof (L1):** spec-audit bỏ qua stale-json (`budget.measured.test-specialist=27723B`) chỉ flag live frontend-designer (26129B) → chứng đo **LIVE** `Get-Item.Length` KHÔNG tin config-cũ (S81 stale-config trap không tái diễn). L2: 10/10 floor + cite-accurate + 27-flag-unchanged. L3: 5/5 caveat + 0 overclaim.
## Deep re-verify (WF3 `wf_6f44eddf-680` — anh-directed 3rd pass, sâu hơn WF2)
3-lane, **KHÔNG chỉ đọc lại** mà inject-lỗi-thật + integration-run:
- **Lane A fault-injection = PASS (răng thật chứng LIVE, KHÔNG reasoning):** 13 inject vào cây tạm (`-RepoRoot`), mỗi cái xác-nhận GAP đúng: A1 bắt `store_memory`/`delete_memory` GAP:HIGH · A1 KHÔNG false-flag body-mention "KHONG store_memory" · A2 bắt xoá-`memoryDelta`/xoá-writeGuard GAP:HIGH · C-spectrum bắt over-cap MED(27008B)/HIGH(32008B>softcap) + **no-false-flag khi có `_INDEX`+gist** + boundary 25600==cap-silent/25601>cap-MED (off-by-one đúng) · B1 bắt detector-missing HIGH/exit≠0 MED · B2 bắt kill-2-anchor LOW · A3 bắt read-only-role-thiếu-decl MED. **8/8 file md5 khớp repo sau restore** (return-only, 0 repo-write). exit-0 mọi lần.
- **Lane C integration = PASS (0 application-defect):** loop-DÒ end-to-end exit-0 + live-read cap (KHÔNG hardcode) + frontend-designer GAP thật (26129B, 0 archive/_INDEX/gist) · **anti-Goodhart proof:** inject `_INDEX`+gist → C-spectrum FLIP GAP→PASS (chứng recompute KHÔNG frozen); mutate cap→99999 → live-echo 99999 + frontend-designer→N/A (chứng single-source live-read) · mfe-eval.ps1 (A2 reuse) exit-0 KHÔNG vỡ · **numbering intact:** §2.1.6→§2.1.7→§2.2 contiguous, §L.b (a)-(h) + MFE-bullet nguyên-vẹn, PHẦN A-I contiguous §I sau §H trước CAVEAT · path-with-space robust.
- **Lane B (cross-doc) = empty-return (#53 fan-out reliability) → em-main self-gate RECOVER:** N=3 nhất-quán (engine §I ↔ ledger) · run-id WF1 khớp 3 chỗ · email hash `54789e19` khớp _index · HCV nhất-quán engine+detector+adap-report · detector **vẫn 27** (0 vocab-fork #4). Consistent trừ 1 gap nhỏ (engine banner WF2 chỉ pointer → **đã fix inline S95**).
### Brittleness Lane A tìm (fail-safe over-flag — NOTE không fix, đúng khuyến-nghị Lane A)
Cả 2 = **fail-toward-flag** (báo DƯ khi format lạ, KHÔNG BAO GIỜ bỏ sót lỗi thật): (1) A2 regex giả-định hmw.js dùng nháy-đơn `'findings'` → nếu reformat nháy-kép sẽ false-GAP (rủi-ro thấp: hmw.js single-writer em-main giữ cú-pháp) · (2) A1 substring-match `store_memory` chưa word-boundary → 1 READ-verb tương-lai chứa chuỗi đó sẽ false-flag (hiện 4 read-verb sạch). Giữ nguyên (over-flag an-toàn hơn miss); ghi honest.
### Dogfood C3 lần 2 (WF3)
WF3-reviewer (Lane A) tự-ghi diary `reviewer/project_s95_*.md` + MEMORY.md pointer (persona vượt return-only) = AS-10 tái-diễn lần-2 → em-main git-diff bắt + **revert** (D9, nhất-quán WF1). cicd-monitor/MEMORY.md = concurrent normal-spawn (KHÔNG return-only) = **giữ** (legit verdict-record). Phân-biệt đúng: return-only-workflow-agent revert vs normal-monitor-spawn keep.
## Honest caveats (broadcast §5 — 5 caveat giữ trong doc §I)
1. **"Vòng chạy thông" = walkthrough THỦ-CÔNG** người-điều-phối nối (đo→phân-loại→ghi), KHÔNG lệnh-tự-động-đơn-nhất (khối-đo exit-0 riêng-lẻ). KHÔNG claim "fully-autonomous".

View File

@ -237,7 +237,7 @@ SE đã có RAG golden-set harness (KHÔNG phải gap): `eval/golden-set-solutio
## PHẦN I — Harness-17: Vòng tự-cải-thiện bộ-nhớ (Đo → Tinh-chỉnh → Kiểm) (🔴 FUNCTION-FLOOR + 🟡 method)
> **Adopt S95 (2026-07-01)** — AI_INFRA 2 broadcast `2026-06-30-Governance-{harness-17-memory-self-improvement, checklist-harness-17}` (hash verify MATCH ✓ cả 2 · `reviewer_gate: PASS`). Áp qua **2-workflow** (Harness-9 mandate): **WF1 `wf_f085d7c4-d78`** (4-agent ground-truth reuse-anchor + draft) + **WF2 review** (run-id → `adap-reports/2026-06-30-Governance-harness-17-memory-self-improvement.md`). Artifact mới: `scripts/memory-selfimprove-audit.ps1` + `.claude/governance/reinject-ledger.md`. **KHÔNG copy luật số ở đây — B1.**
> **Adopt S95 (2026-07-01)** — AI_INFRA 2 broadcast `2026-06-30-Governance-{harness-17-memory-self-improvement, checklist-harness-17}` (hash verify MATCH ✓ cả 2 · `reviewer_gate: PASS`). Áp qua **2-workflow** (Harness-9 mandate) + **1 deep-reverify** (anh-directed): **WF1 `wf_f085d7c4-d78`** (ground-truth reuse-anchor + draft) + **WF2 `wf_8e73c1d6-ee6`** (3-lane review PASS 0-blocking) + **WF3 `wf_6f44eddf-680`** (fault-injection teeth-proof + integration + cross-doc, PASS 0 application-defect). Artifact mới: `scripts/memory-selfimprove-audit.ps1` + `.claude/governance/reinject-ledger.md`. **KHÔNG copy luật số ở đây — B1.**
>
> **Vì sao (đóng "vòng hở"):** nhánh eval (§H MFE) mới ĐO fidelity rồi in số **rồi DỪNG** — thấy mục-sàn rớt khỏi hot-mem nhưng KHÔNG có cơ-chế HÀNH-ĐỘNG. H17 nối 2 khâu thiếu: **Tinh-chỉnh** (tín-hiệu→hành-động, có chốt-dừng) + **Kiểm** (spec-compliance audit). Bộ-nhớ ⇒ **tài-sản được quản-trị**, kiểm-được theo bản đặc-tả.
>