solution-erp/docs/governance/adap-reports/2026-06-29-Governance-harness-16-mfe.md

broadcast: 2026-06-29-Governance-harness-16-{memory-fidelity-eval, mfe-update, cd-judge-sub-update} from: ai_infra project: se (SOLUTION_ERP) date_applied: 2026-06-29 (S93) status: EXECUTED + runtime-smoke (deterministic block) · SCAFFOLD (judge layer) · VERIFIED-pending-restart (command wire) reviewer_gate: PASS (WF2 3-lane: code-gate + cross-harness + honesty, run-id wf_13e3d35a-023)

Adap-report — Harness-16 Memory-Fidelity-EVAL (MFE) → SOLUTION_ERP

TL;DR (nấc honest)

Adopt memory-fidelity-EVAL (MFE) = đo đã-nạp ≠ nhớ-ra ≠ dùng-đúng bằng SỐ ĐO, KHÔNG ước cảm-tính. Deterministic ($0) block = BUILT + WIRED + runtime-smoke S93 (chạy thật, 3-tier, exit 0). Judge layer (recall/apply) = SCAFFOLD-only (seed + stub, chưa scorer — số CHƯA có nghĩa, đúng thiết-kế). Function-floor mục 3-5 của broadcast = ĐẠT; form tự-quyết (SE roster 11-agent). 2-workflow mandate ĐẠT (WF1 implement + WF2 review, run-id bằng-chứng).

Function-floor coverage (broadcast mục 3 — SÀN bắt buộc)

Floor	SE implementation	Nấc
(1) Mẫu-số "bắt-buộc-nhớ" deterministic, status-field KHÔNG grep, dedup	mfe-eval.ps1 LEAD: marks Active(-High) (ACTIVE-MARKS.md status-col) + AS-class + Active-Guards (error-ledger.md Verified-col) + recurring-gotchas (gotchas.md recurrence-token). Dedup deterministic by cross-reference (gotcha/guard mà AS-row trỏ #N/AS-N → gộp; merge_guard_to_as toggle). Hiện = 29 item (4 mark + 13 AS + 5 net-guard + 7 net-gotcha).	✅ executed + WF2 re-derived EXACT
(2a) Coverage% (set có vừa ngân-sách hot không)	byte-span → token-RANGE [bytes/4 .. bytes/3.0] vs cap live-read token_governor.tier1_hotfeed_tokens.lead_tokens (220K). Hiện 29 item ~2.5-3.3K tok = 1.5% worst-case → FIT thừa.	✅
(2b) Effective% (thực-nhớ/dùng — cần trọng-tài)	Judge scaffold (Branch A) — SCAFFOLD-only, số CHƯA có nghĩa (last-resort, đúng broadcast).	🟡 scaffold
(2c) phân-tầng-tuổi (age-band)	FLAG old-but-still-required, KHÔNG cắt (mark RC-…10-29-11). drop chỉ khi status đổi.	✅
(3) Neo Goodhart (đối-chiếu điểm-nhớ vs lỗi-tái-phát thật, KHÔNG self-grade)	in strikes_total=21 + RCA_entries=9 (error-ledger) cạnh coverage; strikes tăng + coverage cao = điểm nói dối (WARN cross-run via .mfe-state.json).	✅
(4) Per-tier + opt-in command param	-Tier lead/sub/all + -Judge. Wire opt-in session-start §2.1.6 + session-end §L.b(c) (param eval, no-param = như cũ). Baseline đầu + retention cuối → chênh = rot trong-phiên.	✅ executed-file (VERIFIED-pending-restart)
(5) Quyết-định 2-ca (thiếu-CHỖ→TĂNG budget / rot→SẮP-XẾP-LẠI)	engine §H3 + README. 🔴 con-số budget = quyền anh (owner-authority); MFE chỉ báo.	✅

Sub-branch (mfe-update + cd-judge-sub-update)

Floor	SE	Nấc
trần đọc-LIVE single-source (không hardcode)	lead_tokens live-read từ budget.json (B1 derived-trỏ-canonical; số đã 60K→200K→220K nên hardcode = rot).	✅ WF2-proven
source-vs-cap = N/A cho sub-tier	SUB-WORKFLOW (ephemeral) = N/A by-design (kênh ký-ức = memory-pack lead bơm).	✅
Branch A judge framework (sample-q + stub scorer + reader, no-API, opt-in, cadence)	eval/mfe/sample-questions.json (12 câu stable-id anchored RC/gotcha#/AS#, immutable, time-stamped — gieo NGAY) + -Judge stub (scorer RỖNG).	🟡 scaffold
Branch B per-role sub-coverage ĐO-THẬT (denom role-file anti-pattern, numer diary, content-word KHÔNG leading-verb)	mfe-eval.ps1 SUB: denom = numbered anti-pattern/NEVER trong agents/<role>.md; numer = diary agent-memory/<role>/MEMORY.md (KHÔNG project/user MEMORY.md — bẫy H15-v3); match = exact-id OR ≥2 content-word sau stop-list. ĐO-THẬT: impl-backend 94% · impl-frontend 85% · reviewer 88%.	✅ executed + WF2 re-derived
bẫy leading-verb (verb-only ăn 9/9 giả)	stop-list (EN+VN verb) + NFD-strip-diacritics → verb-only = 0 content-word = MISS. WF2 adversarial-probed: "NEVER commit push"→0 word→no false-max.	✅ probed
vai prose-only / phù-du = N/A có lý-do (KHÔNG 0%)	den=0 → "N/A (no numbered/NEVER block)" KHÔNG "0%" (div-by-0 guarded). 3 role prose-only (database-agent/harvest/tooling).	✅ (refinement em-main sau smoke)

Tailored (form tự-quyết — SE-specific)

• Denominator = chính governance của SE (4 mark Active-High + 13 AS-class + Active-Guards + recurring-gotchas) — trùng value_protect.patterns là CỐ Ý (cùng tập giá-trị-cao, mark RC-…10-29-11), KHÔNG double-count (dedup by AS cross-reference).
• Vocab-fork resolved (WF1 reviewer bắt — BLOCKING): "memory-fidelity" ĐÃ là H6.7 memoryDelta-routing-fidelity. → đặt tên MFE = memory-fidelity-EVAL ở 4 site, ghi alias-map §H, KHÔNG seed $aliasSets (2-khái-niệm-1-từ, seed sẽ flag-sai). C3-comment chống maintainer seed nhầm.
• READ-ONLY scoped honest: MFE đọc token_governor (single-source) KHÔNG ghi; CÓ ghi 1 file .mfe-state.json (strikes cross-run) — honest_caveats.read_only_scope nói rõ "KHÔNG phải writes-nothing".

2-workflow mandate (Harness-9) — ĐẠT

• WF1 IMPLEMENT wf_4c63e1bd-99e (3 agent: lead-denominator-spec + sub-role-denominator + cross-harness-map) → ground-truth. Lesson: sub-agent #53-truncate (return rỗng) NHƯNG ghi finding xuống diary TRƯỚC → recover-from-disk. AS-10 containment fire: 3 WF1-agent self-write agent-memory (return-only violation) → em-main git-diff bắt + revert (giữ cicd-monitor PE-harvest hợp-lệ). Dogfood: memory-governance hoạt-động đúng giữa lúc adopt memory-eval.
• WF2 REVIEW wf_13e3d35a-023 (3 lane: code-gate + cross-harness + honesty) = PASS 0-blocking. RETURN-ONLY honored (0 tracked write — instruction-fix sau WF1). code-gate re-derived denom 29 to-the-digit + leading-verb-trap adversarial-probed-defeated. honesty: 5 floor (a-e) đủ + live numbers KHỚP §H prose (S81 stale-config trap KHÔNG tái diễn).

Honest caveats (broadcast mục 5 — đưa vào doc, không giấu)

1. Token = RANGE (char/4 ≠ token thật; VN ~3.0-3.5 byte/tok). FIT dùng worst-case bytes/3.0. CHƯA đo bằng tokenizer thật của model → hệ-số còn bất-định; 1.5% nằm xa trần nên kết-luận FIT an-toàn, nhưng đuôi-sổ phình thì theo-dõi.
2. Judge = SCAFFOLD, số ĐẾN SAU. recall/dùng-đúng CHƯA có nghĩa tới khi (a) sample-q đủ-chín-tuổi VÀ (b) trọng-tài độc-lập khác-phiên. Tự-chấm-cùng-phiên = vô-nghĩa. KHÔNG nhận "đã đo độ-nhớ".
3. Age = FLAG không cắt (mark RC-…10-29-11).
4. Deterministic = executed + runtime-smoke, KHÔNG "đã đo retention toàn-hệ". LEAD = coverage/FIT (space), KHÔNG retention.
5. +1 conservative dedup bias (WF2 minor): guard run_in_background map khái-niệm AS-5 nhưng Counters ghi looks-frozen (no AS-5 ref) → giữ net-extra (literal cross-reference, đúng nguyên-tắc deterministic). Tighten = data-fix ledger (defer).

Evidence

• Artifact: scripts/mfe-eval.ps1 (302 dòng, NO-API, ASCII #30, exit 0) · .claude/agent-memory/memory-budget.json:mfe config · eval/mfe/{sample-questions.json,README.md} · engine §H + artifact-row · governance-detectors.ps1 C3-comment · wire 2 cadence · agents/README S93 · .gitignore state-file.
• Smoke: 3-tier exit 0; denom 29/220K=1.5%; SUB impl-backend 94%/impl-frontend 85%/reviewer 88%; 3 prose-role N/A; Goodhart strikes 21/RCA 9.
• 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 → MFE opt-in runtime).

Brutal-honest / phản-biện (§M)

KHÔNG có. Broadcast fit SE (SE có đủ governance-corpus làm denominator + 11-agent roster làm Branch B). 8 broadcast cũ chưa-match-tên (checklist-*/v2.0/h10-detector/h14-adopt) = verify-covered S72/S75/S79/S81 (name-drift, không re-apply). H15-v3-sub-load-channels = covered-S83 + fold-vào-H16 (report riêng).