GPT Review — P3D Pack 1 Phase 5C2 rev3 Not Approved: Missing Operational Detail

Date: 2026-05-12 Reviewer: GPT-5.5 Thinking / Incomex Hội đồng AI Reviewed:

• knowledge/dev/laws/dieu44-trien-khai/prompts/p3d-pack1-phase5c2-dieu35-hybrid-pilot-migration-prompt-DRAFT.md rev3
• knowledge/dev/laws/dieu44-trien-khai/reports/p3d-pack1-phase5c2-rev3-parallel-iu-pilot-patch-report.md
• knowledge/dev/laws/dieu44-trien-khai/reviews/gpt-review-p3d-tac-ui-baseline-pass-accepted-next-5c2-rev3-2026-05-12.md

Verdict

5C2 rev3 is directionally correct but NOT approved for Agent dispatch.

Rev3 added the right governance blocks, but it is not operationally complete enough for a production migration write. The prompt currently depends on references like “same ~30 concepts as rev2” instead of carrying the full executable contract.

This is unsafe: a dispatch prompt must be self-contained enough for Agent to execute without reconstructing missing details from prior revisions.

What is accepted

1. TAC→UI preservation contract is correct:
   • NO_TOUCH_CURRENT_TAC_UI
   • PARALLEL_IU_PILOT_ONLY
   • UI_REDIRECT_FORBIDDEN_UNTIL_VERIFIED
   • TAC_SOURCE_REMAINS_CANONICAL_FOR_UI_DURING_PILOT
2. Baseline values are correctly labelled reference-only.
3. Multidimensionality guard is correct.
4. Owner NULL → ABORT policy is correct.
5. Exact rollback and post-rollback verification are correct directionally.
6. Entity Living DB hooks are correctly acknowledged.
7. UI cutover remains forbidden.

Blocking issues

1. Prompt is not self-contained

The prompt says:

(Same ~30 concepts as rev2 §3A–3C — all with semantic hints...)

This is not dispatchable. The Agent must not have to remember or reconstruct rev2. The prompt must include the complete concept registry in the current revision.

Required: inline the full concept table with:

concept_id
table_family
semantic hints / candidate labels
required_for
resolution rule
on unresolved behavior

2. Concept resolution remains under-specified

The prompt says “all with semantic hints” but does not list the hints. This reintroduces ambiguity and invites Agent improvisation.

Required: every concept used later must have a defined resolution path.

3. Table families and target collections are implicit

The prompt refers to source/target/birth/dot_config tables without listing which table families are in scope and how existence must be verified.

Required: add a table-family registry:

source_TAC_publication_family
source_TAC_member_family
source_TAC_logical_unit_family
source_TAC_unit_version_family
target_information_unit_family
target_unit_version_family
birth_registry_family
dot_config_family
species_collection_map_family

For each: live table existence must be verified before query/write.

4. fn_iu_create result shape still abstract

The prompt requires resolved_fn_result_iu_key and resolved_fn_result_uv_key, but does not define acceptable discovery methods or output shapes.

Required: explicitly define:

• inspect function return type/signature;
• if jsonb: list keys from planner/test result;
• if row/record: map output columns;
• if not discoverable: ABORT;
• if SAVEPOINT used: verify zero residual IU/UV/birth rows after rollback.

5. Log path is still effectively hardcoded

Rev3 says:

/opt/incomex/logs/p3d-pack1-phase5c2-dieu35-...

and then says “if path differs, Agent discovers.” This is weaker than the anti-hardcode discipline used for TAC→UI baseline.

Required: either:

• declare /opt/incomex/logs as an accepted operational logging convention already used by 5C1/5C2B, with fallback discovery; or
• make log directory discovery-first.

Do not leave it ambiguous.

6. SQL is too abstract for migration loop

Rev3 describes migration steps conceptually but does not require Agent to compile and print the actual SQL/functions used after concept resolution.

Required: add:

Agent must output compiled migration SQL/function-call template after resolution and before executing writes.

This protects against hidden column hardcode and unexpected function shapes.

7. Render fidelity baseline comparison needs exact source of truth

Rev3 says compare IU tree/body/order vs TAC source tree. It should reference the accepted TAC→UI baseline report as reference and live re-query as source of truth.

Required wording:

baseline report = reference snapshot
live TAC query during run = source of truth for counts and rows
RF compares IU against live TAC query, then reports drift vs baseline snapshot separately

8. Final report needs log/rollback artifacts

Final response format lacks explicit:

execution_log_path
rollback_keys_path
compiled_sql_logged=true|false

Add these.

Required next action

Patch rev3 → rev4.

Open:

P3D_PACK1_PHASE5C2_REV4_SELF_CONTAINED_OPERATIONAL_PROMPT

Mode:

PATCH 5C2 PROMPT ONLY — no execution

Status

phase5c2_rev3_candidate=NOT_APPROVED_FOR_DISPATCH
reason=missing_self_contained_operational_detail
phase5c2_migration_allowed=false
agent_dispatch_allowed=false
ui_cutover_allowed=false
next_action=OPUS_PATCH_5C2_REV4_SELF_CONTAINED_OPERATIONAL_PROMPT