GPT Review — P3D Pack 1 Phase 4 DRAFT v6 Not Approved; v7 Required

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

• knowledge/dev/laws/dieu44-trien-khai/prompts/p3d-pack1-phase4-governance-vocab-species-prep-implementation-prompt-DRAFT.md rev8 / DRAFT v6
• knowledge/dev/laws/dieu44-trien-khai/reports/p3d-pack1-phase4-implementation-draft-v6-patch-report.md
• knowledge/dev/laws/dieu44-trien-khai/reviews/gpt-review-p3d-pack1-phase4-v5-rolled-back-g7-and-v6-directive-2026-05-11.md

Verdict

DRAFT v6 is NOT approved for dispatch.

Opus correctly fixed the main v5 failure direction, but DRAFT v6 still has issues that violate:

If not certainly right, it is wrong.

Do not send DRAFT v6 to Agent.

What v6 fixed well

1. G3 now checks planner function contract before delta/insert.
2. G7 now calls the 9-argument JSONB-returning planner.
3. G7 derives section/publication type from live dot_config after insert.
4. G7 validates JSONB object and key presence before reading.
5. Section C was correctly identified as needing the same signature fix.

Blocking issues remaining

1. Hash provenance regression

DRAFT v6 reverted Section E to the incorrect formula:

unit_version.content_profile.source_hashes.tac_v1 :=
  sha256( normalized_content_of_corresponding_tac_logical_unit )

This contradicts the prior GPT patch. The correct contract is to preserve the original TAC source hash as provenance, not to imply a newly computed normalized hash.

Required contract:

unit_version.content_profile.source_hashes.tac_v1 := {
  algorithm: "sha256",
  input_recipe: "TAC v1 composite source hash as discovered in Phase 3",
  value: <original tac_unit_version.content_hash>,
  source_table: "tac_unit_version",
  source_id: <source tac_unit_version id>
}

2. Out-of-scope regression: “Do not run any agent”

DRAFT v6 again says:

Do not run any agent.

This is ambiguous/contradictory because the approved prompt itself is intended to be run by the assigned Agent. The correct language is:

Do not dispatch nested/secondary agents from inside this task; this prompt itself may be executed only by the assigned Agent after GPT/User approval.

3. G3 still does not verify the exact function signature sufficiently

G3 checks that arg names include the required names, but it does not verify:

• exact argument count;
• argument types;
• overload ambiguity;
• whether the selected function is the one actually callable by G7.

A function with the same names but wrong types/order could pass. Under the current philosophy, this is not enough.

DRAFT v7 must verify either:

• exact pg_get_function_identity_arguments string equals the production-compatible signature; or
• a robust pg_proc check that confirms argument count, names, and types exactly.

Expected identity signature:

p_canonical_address text, p_title text, p_body text, p_actor text, p_unit_kind text, p_section_type text, p_owner_ref text, p_publication_type text, p_parent_ref uuid

Return type must remain jsonb.

4. Section C post-commit probe is useful but must be covered by the same function-contract certainty

Opus was right to fix Section C, but because G3 is still not exact enough, Section C inherits the risk.

Decision

Patch DRAFT v7, narrowly:

1. Fix Section E hash provenance contract.
2. Fix out-of-scope agent wording.
3. Strengthen G3 contract gate to exact planner signature + return type.
4. Keep G7 and Section C 9-arg JSONB logic unless patching G3 requires a matching minor wording update.
5. Do not change delta/insert/rollback/species/hash execution scope.

Status

phase4_draft_v6=REJECTED_FOR_DISPATCH
phase4_draft_v7_patch_allowed=true
agent_dispatch_allowed=false
seed_execution_allowed=false
migration_allowed=false
next_action=OPUS_PATCH_DRAFT_V7_EXACT_G3_AND_RESTORE_GPT_PATCHES