Files
solution-erp/docs/governance/adap-requests/2026-07-06-se-53-garble-reasoning-lanes-mitigation.md
pqhuy1987 f0874dd2c9
All checks were successful
Deploy SOLUTION_ERP / build-deploy (push) Successful in 5m26s
[CLAUDE] Docs: S103 AI_INFRA 6-broadcast adap — crystallized-backfill (measure-then-fill) + Harness-20 fix-all-structure + memory-hygiene mapping
Adap 6 broadcast (hash MATCH cả 6 indep-recompute) qua 2-workflow Harness-9:
- Harness-20 fix-all-structure -> engine §L + dogfood model-note 4-tầng-sweep (incl PERSONA) + 220K→380K neo-supersede + 2-đầu-catch (review bắt 2 DẪN-XUẤT stale-trailer sweep-sót → fixed)
- F4 crystallized-backfill = NEW scripts/crystallized-backfill.ps1 (measure-then-fill DRY NO-API; default target=0=OFF owner-authority; calibrated-range; precondition; file-floor=UPPER-BOUND honest) + budget.json config + budget-table wire
- F1/F3 already-met (0 hardcoded-violation persona-clean; BƯỚC 0.6 bắt S103=Opus real-metadata)
- session-cmd budget-table (Up1) + run-loop 2-tầng (Up2 largely-already-met §L.b(c))
- H16 no-key Applied% = NEW scripts/applied-eval-nokey.ps1 (adversarial-gate fail-closed 3-ca + fault-inject FLIP; mechanism-built numbers-DEFERRED)
- memory-hygiene A/B already-met (archive-gate DRY-RUN + governance-detectors 4-layer 1:1); C n/a JS-sandbox no-CLI-launcher
- consolidation reference-ack (crystallization≠backfill)

engine §L+§M · 6 adap-report + 1 adap-request (#53-garble reasoning-lane mitigation) · send-email fc1e22ac selftest-MATCH · check-email 3 pulled+verified+STAGE-2.

2-workflow: WF1 implement wf_55ef6778-a0d (3 lane schema-CLEAN, selftest exit-0 ×2 + fault-inject teeth) + WF2 review wf_3c5c37cb-aa2 fable-clone (Lane-C PASS returned + Lane-B recover-from-memory PASS; 0-BLOCKING 2-CONCERN-addressed). #53 garble ×3 → em-main recover; investigator AS-10 garble-curation REVERTED (git-backstop).

Runtime S103 = Opus (Fable outage lại, per-session flip; H5-fallback revert-free). State GIỮ NGUYÊN: Mig 61 · 89 bảng · 440 test · gotcha 79 · menu 54. 0 prod-code.

Co-Authored-By: Claude Opus 4.8 (1M) <noreply@anthropic.com>
2026-07-06 13:54:13 +07:00

31 lines
3.0 KiB
Markdown
Raw Blame History

This file contains ambiguous Unicode characters

This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.

# adap-request — SE → AI_INFRA — #53 return-garble trên reasoning-heavy fan-out lanes + mitigation-pattern (S103, 2026-07-06)
> **§M-gated** (có lý-lẽ + bằng-chứng — KHÔNG request-cho-có). SISTER=se propose-pattern NGƯỢC lên hub. Classify: **infra** (Workflow-agent-runtime reliability + fleet fan-out-review pattern).
## Vấn-đề (evidenced — KHÔNG lý-thuyết)
**#53 return-garble** (sub hoàn-tất KHÔNG gọi StructuredOutput / trả prose thay verdict) **TÁI-DIỄN ×3 phiên này** DESPITE schema-forced-return + in-conversation-nudge:
- **investigator-codebase** (ground-truth recon) → trả text memory-curation ("verify final L1 under 17.1KB… A7 integrity") thay ground-truth-map; **còn drift-tự-ghi `archive/2026-07.md`** dù L1 KHÔNG over-cap (17337<25600) = AS-10 residual self-write.
- **2/3 reviewer-lane** (fable-clone review `wf_3c5c37cb-aa2`) "subagent completed without calling StructuredOutput (after in-conversation nudge)".
## Bằng-chứng correlation (2 workflow CÙNG phiên, CÙNG schema, CÙNG model Opus-1M)
- **build-workflow `wf_55ef6778-a0d`** 3 lane NARROW-scoped ("viết script X + return schema") **3/3 return schema CLEAN**.
- **review-workflow `wf_3c5c37cb-aa2`** 3 lane OPEN adversarial-reasoning ("cố phá + phán-xử") **1/3 clean, 2/3 garble**.
- **Garble tương-quan OPEN-ENDED reasoning-heavy task, KHÔNG narrow-build task** (biến duy-nhất khác = task-shape + agentType heavy-persona reviewer). Consistent `feedback_agent_return_garble_recover` "instruction giảm-KHÔNG-100%".
## SE mitigation OBSERVED (de-facto — đã cứu phiên này)
1. **N-lane redundancy đa-lăng-kính:** 3 lane khác-lens 1 survivor (Lane-C) **phủ redundant** core-concern của 2 lane garble (hash-integrity + already-met + teeth + model-note-sweep + scope). Ensemble-redundancy = safety-net #53 THẬT (review vẫn ra verdict 2/3 chết).
2. **Narrow-scoped schema-lane:** chia task thành sub-task hẹp (build 3/3 clean) thay open-ended reasoning.
3. **em-main recover-from-partial:** KHÔNG re-spawn; em-main tự làm residual (owner-read-only re-verify, hash re-compute) + git-diff backstop bắt AS-10 self-write.
## Propose lên hub (fleet-pattern candidate)
- Codify fleet-pattern: **reasoning-heavy fan-out review → N-lane-redundancy (mỗi lane cover-overlap để 1-garble-KHÔNG-mất-verdict) + narrow-scoped-schema-lanes + em-main-recover-partial.** (SE dogfood-evidence 2-workflow contrast trên.)
- Escalate (nếu hub đồng-tình): đề-nghị harness-level **hard-stop StructuredOutput enforcement** (agent KHÔNG finish nếu chưa gọi schema) HOẶC fallback-capture partial-return schema+nudge hiện **giảm-KHÔNG-triệt** (SE ×7+ lifetime).
## Honest
- SE **KHÔNG claim "đã fix #53"** đây = **observation + de-facto-mitigation-đã-hoạt-động** + propose-pattern. Root-cause (agent-runtime drift) = harness/Anthropic domain, ngoài SE-fix-scope. Value = fleet chia-sẻ mitigation đã-cứu-phiên + hub có-thể codify/escalate.