GPT Review — 20B Closure and Return to Birth Pipeline

Date: 2026-05-05 Reviewer: GPT-5.5 Thinking / Incomex Hội đồng AI Reviewed:

• knowledge/dev/laws/dieu44-trien-khai/reports/20b-p4-context-pack-filesystem-retention-report.md
• knowledge/dev/laws/dieu44-trien-khai/reviews/opus-review-20b-p4-report-and-20b-closure-2026-05-05.md
• knowledge/dev/laws/dieu44-trien-khai/reports/p2b-inv-birth-pipeline-inventory-report.md

Verdict

20B lifecycle COMPLETE. Vector/search emergency is resolved enough to return to the paused birth/IU work.

Accepted 20B results

• 20B-P1: context-pack KB upload disabled via reversible flag.
• 20B-P2: 1,175 KB context-pack docs inventoried.
• 20B-P3: 1,175/1,175 KB context-pack docs purged; context-pack/ count = 0; search regression 0/35 context-pack hits.
• 20B-P4: filesystem retention wrapper + cron installed; initial cleanup removed 79 old dirs; current symlink target preserved.

The direct cause of vector pollution is fixed:

• no new context-pack KB mirror should be created by default;
• old context-pack KB mirror is gone;
• filesystem archive is bounded by retention;
• search now surfaces canonical docs again.

Remaining vector hygiene work

These are not blockers for returning to Đ44/IU, but should become TDs:

1. Verify cron after one scheduled run:
   • tail /var/log/incomex/context-pack-retention.log
   • confirm no current-target deletion.
2. Đ43/20C semantics:
   • kb_mirror_status still may say live even when KB upload is disabled.
   • Needs formal semantics patch later.
3. KB Governance Framework:
   • doc tiers, TTL, upload metadata, quota alerts, search regression suite.
4. Optional 20A search filter as defense-in-depth.

Return to paused work: birth/IU pipeline

Before resuming Pack 2C or designing dot-iu-create, process the birth inventory report first.

The report changed the understanding substantially:

• Runtime birth pipeline has 162 birth trigger instances, not 31.
• QT-002 wrapper tool is not runtime-active; it is policy/design, while actual runtime safety belt is PG trigger fn_birth_registry_auto.
• QT-001 backfill exists as dot-birth-backfill but has no cron.
• Orphan/ghost detection exists via fn_registry_health() and multiple scanners, but direct warning when backfill/auxiliary engine runs is missing.
• Several docs are stale against runtime: trigger count, birth_registry schema, duplicate DOT registrations, registry entries without binaries.

Next directive to Opus

Create a birth pipeline reconciliation note before more IU/Pack 2C work:

knowledge/dev/laws/dieu44-trien-khai/reviews/opus-birth-pipeline-inventory-review-and-reconciliation-2026-05-05.md

Required sections:

1. Confirm P2B-INV report PASS / any gaps.
2. Reconcile User philosophy with runtime reality:
   • main channel / canonical path;
   • safety-belt trigger;
   • auxiliary/backfill channel;
   • orphan detector;
   • ghost detector;
   • warning gap.
3. Separate:
   • what already exists and runs;
   • what exists only in docs/design;
   • what is stale doc;
   • what is hardcoded legacy code.
4. Explicitly state: do not redesign birth process; hardcode/stale docs are bugs or TDs under existing process.
5. Recommend next execution/design order:
   • likely first: patch docs/registry truth about runtime counts and schema, or create dot-iu-create wrapper design under existing pipeline;
   • do not jump into new lifecycle design.
6. If any missing evidence remains, propose one narrow read-only prompt only.

Hard boundaries for next Opus step

• no runtime mutation;
• no new birth process design;
• no Pack 2C dispatch yet;
• no DOT patch yet;
• no cleanup/delete;
• just reconcile and propose next order.

Current state

Vector/search is clean enough to resume. The next immediate work is conceptual/runtime reconciliation of birth pipeline inventory, then decide whether to proceed with IU wrapper/Pack 2C or documentation/health-check fixes first.