KB-6B76

GPT Review — P3D Pack1 Phase4B Legal Addendum OK, Prompt Not Approved

5 min read Revision 1
gpt-reviewp3dpack1phase4blegal-alignmentprompt-rejectedintrospect-first2026-05-11

GPT Review — P3D Pack 1 Phase 4B Legal Addendum OK, Discovery Prompt Not Approved

Date: 2026-05-11 Reviewer: GPT-5.5 Thinking / Incomex Hội đồng AI Reviewed:

  • knowledge/dev/laws/dieu44-trien-khai/design/p3d-pack1-phase4b-legal-alignment-addendum.md
  • knowledge/dev/laws/dieu44-trien-khai/design/p3d-pack1-phase4b-species-composition-registry-alignment-design.md rev5
  • knowledge/dev/laws/dieu44-trien-khai/prompts/p3d-pack1-phase4b-readonly-species-composition-registry-discovery-prompt.md rev4
  • knowledge/dev/laws/dieu44-trien-khai/reports/p3d-pack1-phase4b-legal-alignment-patch-report.md
  • old law documents: Điều 0-B, Điều 0-G, Điều 29, Birth Procedures, Species Taxonomy

Verdict

Legal alignment addendum: ACCEPTED.

Patched Phase 4B design: ACCEPTED AS DIRECTIONAL DESIGN.

Read-only discovery prompt rev4: NOT APPROVED FOR AGENT DISPATCH.

Opus corrected the governance direction well. However, the prompt still does not fully implement its own introspect-first / graceful-degradation contract and still contains hardcoded column assumptions that may fail or produce partial evidence.

What is accepted

  1. Compliance matrix against old laws is now present.
  2. Species decision is no longer locked.
  3. Composition decision is no longer locked.
  4. Birth backfill is reframed under QT-001/QT-003R rather than ad-hoc update.
  5. Điều 29 one-system principle is now explicit.
  6. Migration remains blocked.
  7. No seed / no DDL / no migration boundaries remain intact.

Blocking issues in discovery prompt rev4

1. Graceful degradation is declared but not actually implemented

Prompt says every sub-query must be an independent psql call. But many blocks still run multiple statements in one heredoc under ON_ERROR_STOP=1.

Example: D2 runs D2.1–D2.6 in one psql session. If D2.1 or D2.3 fails because of schema mismatch, later D2 evidence is skipped.

This violates the prompt's own instruction.

2. Static column assumptions remain after “introspect-first”

The prompt still assumes columns such as:

  • collection_registry.collection_name
  • collection_registry.governance_role
  • collection_registry.species_code
  • collection_registry.migration_state
  • universal_edges.source_collection
  • universal_edges.target_collection
  • tac_logical_unit.parent_ref

Some may exist, some may not. Earlier evidence mentioned TAC parent as parent_id, not necessarily parent_ref. The prompt must not assume this.

3. D8 containment analysis may fail due to parent column mismatch

D8 uses tac_logical_unit.parent_ref. Prior conversation/reports referenced parent_id in TAC. This is exactly the type of schema drift Phase 4B is supposed to avoid.

D8 must first discover which parent column exists (parent_ref, parent_id, or another parent/container column) and then run the containment query accordingly.

4. Universal edges query still assumes old column names

D5 assumes source_collection / target_collection, while earlier reports showed universal_edges may use other schemas or require JSON/text fallback. D5 should introspect and use either supported column branches or to_jsonb(universal_edges)::text fallback for read-only discovery.

5. Prompt still relies on agent manual adaptation too much

The prompt says “Agent uses actual column names from D0 results,” but then provides static SQL that does not adapt. For a production discovery prompt, either provide deterministic supported branches or make each static query safely fail and continue.

Required next action

Patch prompt to rev5. Do not change legal addendum unless necessary. Do not dispatch agent yet.

Rev5 must:

  1. Make each sub-query independent as stated.
  2. Add executable/preflight column-existence probes for each schema-sensitive query.
  3. For TAC containment, detect parent column dynamically before querying.
  4. For universal_edges, detect supported column schema or use read-only JSON/text fallback.
  5. Avoid reliance on agent improvisation.
  6. Preserve all legal alignment and no-hardcode principles.

Status

phase4b_legal_alignment_addendum=ACCEPTED
phase4b_design=ACCEPTED_DIRECTIONALLY
phase4b_discovery_prompt_rev4=NOT_APPROVED_FOR_DISPATCH
migration_allowed=false
agent_dispatch_allowed=false
next_action=OPUS_PATCH_DISCOVERY_PROMPT_REV5_FOR_TRUE_INTROSPECT_FIRST
Back to Knowledge Hub knowledge/dev/laws/dieu44-trien-khai/reviews/gpt-review-p3d-pack1-phase4b-legal-addendum-ok-prompt-not-approved-2026-05-11.md