GPT Review — VRC Report + Directive to Opus

Date: 2026-05-02 Source reviewed: knowledge/dev/laws/dieu38-trien-khai/reports/vector-reality-check-agent-data-qdrant-2026-05-02.md rev 1 Status: VRC ACCEPTED; legacy vector repair must be separated from IU-0 design

Verdict

Codex VRC report is accepted as sufficient evidence for IU-0 outline revision. It answered the required VRC questions with runtime evidence and no infrastructure mutation, except report creation.

Opus evaluation is directionally correct, but must be adjusted in one important governance point: fixing V1/V2 is urgent, but must not be mixed into IU-0 or done casually. It must be handled as a separate Legacy Vector Stabilization track with APR/runbook/rollback.

Critical findings accepted

1. Legacy vector is document-based, not unit-based.
2. Current Qdrant collection is production_documents, 1536-dim cosine, 11052 points.
3. Current payload has document_id, content, metadata.chunk_index, metadata.total_chunks; no unit_id, canonical_address, or unit_version_id.
4. Search dedups only by document_id; no semantic dedup, no rerank, no multi-collection search.
5. Duplicate KB draft + IU final will likely return both results unless search adapter/dedup is designed.
6. Same-collection payload tagging is not safe as Phase 0 because current search filters only tags/status.
7. Safest IU direction: Phase 0 read-only mapping adapter; Phase 1 separate IU collection; Phase 2 search adapter/dedup; Phase 3 unit-aware chunking.
8. Legacy system has active defects: missing PG trigger, broken scheduled dot-vector-audit, 47 orphan vectors, 7 ghost docs, health ratio critical, unclear Qdrant backup schedule.

Strategic decision

Two tracks must be separated:

Track A — Legacy Vector Stabilization

Urgent operational repair for existing KB vector health. This is not IU-0 and not P44-6.

Scope for next Opus output:

• create an OUTLINE/RUNBOOK only, not execute;
• name it LV-0 or TD-VECTOR-STAB-0;
• include V1/V2/V3/V4/V5 evidence;
• propose read-only verification first;
• propose guarded repair steps for missing trigger and failed audit schedule;
• include rollback plan and no-bulk-reembed rule;
• do not run cleanup, reindex, trigger creation, or service restart without explicit approval.

Track B — IU-0 OUTLINE-C

Revise IU-0 vector section based on VRC evidence.

Scope:

• document-based legacy remains untouched;
• IU Vector is parallel;
• Phase 0 mapping adapter only;
• Phase 1 separate IU collection preferred;
• Phase 2 search adapter with duplicate resolution;
• Phase 3 unit-aware chunking after benchmark;
• duplicate content is first-class critical/high risk.

Directive to Opus

Do not ask Codex for more investigation now. The VRC is sufficient for outline-level work.

Next deliverables, in order:

1. Draft Legacy Vector Stabilization OUTLINE (LV-0) — no execution.
2. Draft IU-0 OUTLINE-C revised from OUTLINE-B using VRC evidence — no upload, no full draft.
3. Return both as concise artifacts/reports for GPT/User review.

Non-negotiable guardrails

• Do not modify Qdrant, PG, triggers, cron, containers, code, or vector data.
• Do not call /kb/audit-sync with auto_heal=true.
• Do not call cleanup/reindex endpoints.
• Do not create trigger or fix cron yet.
• Do not re-embed or bulk refresh vectors.
• Do not open P44-6.
• Do not upload IU-0.
• Do not treat IU collection as approved implementation; it is design-only until APR.

Placement guidance

• VRC report stays in knowledge/dev/laws/dieu38-trien-khai/reports/.
• IU-0 design outline stays under knowledge/dev/laws/dieu44-trien-khai/design/07-iu0-information-unit-minimum-standard-outline.md if uploaded later.
• Legacy Vector Stabilization report should go under knowledge/current-state/reports/ or knowledge/dev/laws/dieu38-trien-khai/reports/ only after User/GPT approves upload; for now, artifact only.

Review conclusion

Proceed with controlled design work. Do not execute fixes yet. Legacy vector repair is urgent, but must be isolated, reversible, and approved separately from IU-0.