GPT Review — P3D Pack 1 Phase 4 Implementation DRAFT v2 No-Hardcode Fail

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 rev2
• knowledge/dev/laws/dieu44-trien-khai/reports/p3d-pack1-phase4-implementation-prep-prompt-patch-report.md
• knowledge/dev/laws/dieu44-trien-khai/reports/p3d-pack1-phase4-governance-vocab-discovery-report.md

Verdict

NOT APPROVED FOR DISPATCH.

Opus improved the prompt substantially, but DRAFT v2 still contains hardcode leakage and SQL/schema risks. Under the no-hardcode directive, this cannot be sent to Agent.

What improved

• Vocab values are now derived from live TAC data and IU vocab deltas.
• The fixed manual 14-row insert list is gone.
• Temp-table delta pattern is the right direction.
• Migration remains prohibited.
• DDL remains prohibited.

Blocking issues

1. Invalid dot_config column assumption

Discovery report says dot_config has:

key, value, description, updated_at

DRAFT v2 inserts into:

dot_config (key, value, description, created_at)

This is wrong and would fail or reveal that the prompt is not schema-driven.

2. Species mapping still hardcodes target rows

DRAFT v2 still inserts explicitly:

'information_unit'
'unit_version'

This violates the decision to derive target collections from live collection_registry. Stable table names may be used for a scoped DDL pack, but seed execution must use live-discovered target collection rows.

3. Species composition hardcode remains

DRAFT v2 inserts:

'unit_version', es.species_code, 'subordinate'

This is hardcoded composition. The previous directive explicitly said to use discovered species/composition fields and STOP if the mapping policy is not uniquely derivable.

4. Species SQL is likely invalid

This pattern is unsafe/invalid:

SELECT 'information_unit', es.species_code, es.composition_level
FROM entity_species es
...
HAVING count(*) = 1

It mixes non-aggregated selected columns with aggregate HAVING and no GROUP BY / scalar CTE. The uniqueness check should be a CTE, not a HAVING after raw rows.

5. Rollback by description prefix is unsafe

DRAFT v2 proposes:

DELETE FROM dot_config WHERE description LIKE 'Seeded from TAC live discovery%';

This can delete rows from previous legitimate runs. Rollback must delete only rows inserted by this run, identified by a unique run marker or exact captured inserted keys.

6. ON CONFLICT assumptions are not verified

The prompt assumes dot_config has a unique conflict target on key and species_collection_map has usable conflict constraints. It must introspect indexes/constraints before using ON CONFLICT.

7. STOP if delta ≠ expected is ambiguous and can reintroduce hardcode

There must be no expected count/list from report memory. The only allowed checks are schema compatibility, conflict safety, uniqueness of law species, and whether computed delta is empty/non-empty.

8. law_unit is still typed as a fixed seed value without a derivation rule

law_unit can be a required migration contract value, but the prompt must state its source: active Pack 1 migration contract requiring law units, not a discovered TAC column. It must be treated as one contract-required value, not a live table count.

Required next action

Opus must patch DRAFT v2 to DRAFT v3 or GPT may patch directly. DRAFT v3 must be split conceptually into:

1. Vocab prep — likely executable after GPT review.
2. Species mapping prep — either executable only if fully schema/uniqueness-safe, or deliberately deferred if not safe.
3. Hash provenance contract — document JSONB contract only, no backfill.

DRAFT v3 required properties

• Introspect dot_config columns and constraints before seed.
• Insert only into existing columns.
• Generate a unique run marker.
• Capture inserted rows in temp tables.
• Rollback only exact rows inserted by this run.
• Compute vocab deltas from live data.
• Treat law_unit as contract-required and clearly label it.
• Do not hardcode species mapping rows.
• Do not hardcode subordinate composition.
• Use CTE uniqueness checks for law species.
• If species mapping cannot be derived safely, defer it rather than pretending it is safe.
• No migration, no DDL, no Directus, no Qdrant.

Status

phase4_implementation_draft_v2=REJECTED_FOR_DISPATCH
reason=NO_HARDCODE_AND_SCHEMA_SAFETY_FAILURES
migration_allowed=false
seed_allowed=false_until_DRAFT_v3_GPT_REVIEW
next_action=PATCH_DRAFT_V3_WITH_SCHEMA_INTROSPECTION_RUN_MARKER_AND_NO_HARDCODE_SPECIES_POLICY