Files
solution-erp/docs/governance/adap-reports/2026-06-30-Governance-harness-17-memory-self-improvement.md
pqhuy1987 18a17f5498
All checks were successful
Deploy SOLUTION_ERP / build-deploy (push) Successful in 5m8s
[CLAUDE] Docs: adopt Harness-17 memory self-improvement loop (Đo→Tinh-chỉnh→Kiểm) — 2-workflow, 0 prod code
- engine §I codify 10 red-floor: A1 HCV/A2 load-fid/B1 signal→action/B2 CG-1 termination/B3 build-gap-vs-floor-rot/C1 spec-audit/C2 spot-check/C3 single-writer/C4 provenance/CG-2 age-exempt
- REUSE-cite (no rebuild): A2=§H MFE · C3=D9 · CG-2=§G.2b value_protect · archive=§A
- new mechanized: scripts/memory-selfimprove-audit.ps1 (C1 spec-audit NO-API exit-0, runtime-verified — bắt frontend-designer over-cap GAP thật)
- new: .claude/governance/reinject-ledger.md (B2 CG-1 git-tracked append-only N=3, tách khỏi budget.json)
- wire session-start §2.1.7 (loop-DÒ) + session-end §L.b(c) (loop-REFINE)
- vocab: HCV alias no-seed (coverage 3-way) + governance-detectors C3-comment; cap-fix 220K→380K-pointer
- 2-workflow (Harness-9): WF1 wf_f085d7c4-d78 (ground-truth) + WF2 wf_8e73c1d6-ee6 (3-lane PASS 0-blocking)
- adap-report + checklist-pointer + email AI_INFRA (hash 54789e19, hash-verified self-consistent)
- dogfood: WF1 agent self-write MEMORY.md reverted (D9/AS-10 containment; archive missing → saved 8.5KB)

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
2026-07-01 09:11:46 +07:00

9.2 KiB
Raw Blame History

broadcast: 2026-06-30-Governance-{harness-17-memory-self-improvement, checklist-harness-17} 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) hash_verify: MATCH both broadcasts (content_sha256 recompute = declared; PowerShell canonical)

Adap-report — Harness-17 Vòng tự-cải-thiện bộ-nhớ (Đo → Tinh-chỉnh → Kiểm) → SOLUTION_ERP

TL;DR (nấc honest)

Adopt Harness-17 = memory self-improvement loop (Eval → Refine → Audit) — đóng "vòng hở" của nhánh eval (§H MFE): trước chỉ ĐO fidelity rồi in số rồi dừng; H17 nối Tinh-chỉnh (tín-hiệu→hành-động có chốt-dừng) + Kiểm (spec-compliance audit). 10 mục 🔴 sàn ĐẠT (WF2-L2 verified 10/10). Nhiều mục = REUSE-cite cơ-chế SE sẵn-có (A2=§H MFE · C3=D9 · CG-2=value_protect · archive=§A) — KHÔNG rebuild. Mục MỚI: C1 spec-audit script (memory-selfimprove-audit.ps1, mechanized, runtime-verified) + B2 reinject-ledger + B1 signal→action map + B3 build-gap-vs-floor-rot test. 2-workflow mandate ĐẠT (WF1 implement wf_f085d7c4-d78 + WF2 review wf_8e73c1d6-ee6). 0 production code — governance-only. Honest nấc: phần lớn B/C = CONVENTION (đúng như checklist H17 tự nói); chỉ C1-checker + A2-MFE + strike-anchor + HCV-proxy = mechanized.

Function-floor coverage (10 mục 🔴 — WF2-L2 verified 10/10)

Floor SE implementation Nấc
A1 HCV (harvest-coverage, said-vs-kept) harvest-curator H2 trục Coverage (convention) + §H MFE SUB per-role (mechanized) + HCV proxy MỚI trong memory-selfimprove-audit.ps1 (run-trace sub-*.md có synthesis = file-exists). Đặt tên HCV tránh "coverage" 3-nghĩa (alias-map §I). 🟡 N/A-idle (proxy sẵn-sàng, chưa có run-trace giữ sub-*.md để chấm; ratio = calibration-interim)
A2 load-fidelity (fit + %/tier) = §H MFE (mfe-eval.ps1 LEAD Coverage-FIT + SUB per-role). REUSE-cite. (S93, runtime-smoke)
B1 signal→action map (4 loại) bảng §I: floor-rot→reinject-verbatim · repeat≥2→promote · old+valueless→cold-archive · raw→distill (/sleep-recovery-memory-l2). Mỗi hàng map cơ-chế SE THẬT. executed-file (detect=mechanized · action=convention)
B2 CG-1 termination-ledger .claude/governance/reinject-ledger.md (git-tracked · append-only · N=3 · escalate-anh). TÁCH khỏi budget.json (§G.4(4) owner-authority). executed-file (convention)
B3 build-gap vs floor-rot phép-thử-ranh-giới §I: "item ĐÃ-từng-tồn-tại (git-history) ∩ CÒN-giá-trị ∩ RỚT-L1?" → floor-rot=reinject; chưa-dựng=build-gap. executed-file (convention)
C1 spec-audit LIVE scripts/memory-selfimprove-audit.ps1 (NO-API, ASCII #30, exit-0, emit CRITERIA+GAP KHÔNG frozen-tally). 3 mặt: (a) write-gov [tool-scope+hmw.js schema+propose-only] · (b) change-gov [COMPOSE governance-detectors] · (c) distillation-spectrum [over-cap có _INDEX/gist]. executed + runtime-verified S95 (bắt frontend-designer over-cap GAP thật)
C2 3 spot-check người harvest-fidelity + verdict + provenance = harvest-curator 5-trục + reviewer + session-end §L.b(f). convention
C3 single-writer (loop propose-only) = D9 (§D.3). REUSE-cite. (honest: xem §M)
C4 provenance 2-chiều spawn-record hmw.js:53-63 + session-end §L.b(d). 🟡 partial (nhãn ngầm; field-set lệch — xem §M)
CG-2 high-value miễn-tuổi = §G.2(b) value_protect + mark RC-…10-29-11. REUSE-cite. convention

Tailored (form tự-quyết — SE-specific)

  • Vocab (C3 — WF1 Agent-D bắt): "coverage" nay 3 nghĩa (H2 harvest "Coverage" 0-silent-miss · H16 MFE SUB "coverage" role-floor-retention · H17 A1 "HCV" said-vs-kept). → đặt tên HCV, alias-map §I, KHÔNG seed $aliasSets (giống H16 memory-fidelity; 3-khái-niệm-1-từ seed sẽ false-flag). C3-comment governance-detectors.ps1. "audit"→"spec-audit". Từ mới reinject/floor-rot/build-gap/signal→action = grep 0-match → sạch. vocab-fork giữ nguyên 27 flag (WF2-L2 verified, KHÔNG tạo fork thứ 4).
  • N=3 (SE solo-dev): reinject ≤1 lần/3 phiên (nhỏ hơn multi-dev vì solo = feedback-loop nhanh).
  • Cadence wire: session-start §2.1.7 (loop-DÒ INFORM-only) + session-end §L.b(c) (loop-REFINE propose). Opt-in eval (tương-thích-ngược).
  • 0 hard-conflict (WF1 Agent-D + WF2-L2): 3 placement-constraint tuân đủ — reinject-ledger tách budget.json · reinject=promote-COPY giữ D11 · loop=D6-under-D9 (CG-1 escalate ≙ E.2).

2-workflow mandate (Harness-9 / adap-apply §31) — ĐẠT

  • 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.

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".
  2. Judge recall/apply = STUB (mfe-eval -Judge scaffold) — số nhớ/dùng CHƯA nghĩa tới khi sample-q chín-tuổi + trọng-tài độc-lập khác-phiên. Tự-chấm-cùng-phiên = vô-nghĩa.
  3. HCV threshold = calibration-interim (cò-báo-đuôi, KHÔNG vạch-chất-lượng chốt). HCV proxy hiện N/A-idle (chưa run-trace giữ sub-*.md).
  4. Tự-soi (meta-blind-spot): vòng đo agent-công-nhân, CHƯA đo chính các khâu-meta (đo/tinh-chỉnh/kiểm tự-nó).
  5. Cô-đặc = compress-only; verify-trước-khi-giữ CHƯA dựng.

Evidence

  • Artifact MỚI: scripts/memory-selfimprove-audit.ps1 (19KB, NO-API, exit-0, runtime-verified) · .claude/governance/reinject-ledger.md (4KB, B2).
  • Modified: docs/governance/harness-11-engine.md §I (10-floor codify + artifact-map row + fix stale cap-comment 220K→380K-pointer) · .claude/commands/session-start.md §2.1.7 · .claude/commands/session-end.md §L.b(c) · scripts/governance-detectors.ps1 C3-comment HCV-no-seed.
  • Runtime: spec-audit exit-0, CRITERIA A1 A2 A3 B1 B2 C-spectrum:frontend-designer, GAP C-spectrum:frontend-designer:MED (răng thật). governance-detectors vẫn 27-flag (no vocab-fork thứ 4).
  • 0 production code. State THẬT GIỮ NGUYÊN (Mig 59 · 88 bảng · 434 test · gotcha 76 · bundle admin CDn40vpV/user CoPmPBg7).
  • ⚠️ restart CLI (command .md no hot-reload → §2.1.7 + §L.b(c) wire runtime).

Brutal-honest / phản-biện (§M)

  • C3 store_memory-strip honest-correction: broadcast + D9 nói "store_memory strip mọi sub". WF1 Agent-A phát-hiện: strip = hook AI_INFRA runtime, KHÔNG có file .js trong repo SE — SE enforce tại-chỗ = tool-scope (11 sub tools: chỉ READ verb, spec-audit A1 verify) + hmw.js SCHEMA return-delta-only (A2 verify). Đã sửa overclaim tiềm-ẩn ở §I + spec-audit self-disclosure (D9 wired-not-hook; .claude/settings.json KHÔNG có hook → AS-10 fails-open).
  • C4 field-set LỆCH: hmw.js {task·verdict·learned·surprise} (runtime-return) ≠ session-end {agent·task·nấc·evidence} (harvest-record) — 2 schema 2 tầng khác nhau, KHÔNG assume đồng-nhất. provenance-2way-tag tường-minh = build-thêm nhỏ (defer, honest — chưa làm).
  • A1 HCV nấc thật: deterministic proxy = file-exists (run-trace synthesis); said-vs-kept RATIO = calibration-interim + hiện N/A-idle (chưa mẫu-chấm). KHÔNG claim "đã đo said-vs-kept".
  • Double-run minor (WF2-L1 note): session governance chạy governance-detectors 2× (§2.1.3 direct + §2.1.7 spec-audit child-invoke). Đã disclosed caveat-1; detector = fast NO-API grep; redundancy cố-ý (composition). 0 correctness impact.
  • KHÔNG phản-biện function-floor: H17 fit SE (SE có đủ eval-branch §H MFE + archive-gate §A + D9 + value_protect làm nền → đóng-vòng tự-nhiên). Broadcast honest-tier khớp SE honest-tier (phần lớn convention, mechanized rõ đâu).