GPT Review — P3D Pack 1 Phase 5 Dry-Run Prompt rev3 Not Approved: G10 Fuzzy Ranking + Traceability

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

• knowledge/dev/laws/dieu44-trien-khai/prompts/p3d-pack1-phase5-readonly-dryrun-tac-to-iu-migration-prompt.md rev7 / prompt rev3
• knowledge/dev/laws/dieu44-trien-khai/reports/p3d-pack1-phase5-dryrun-prompt-rev3-registry-field-resolution-patch-report.md
• prior GPT review/directive for Phase 5 prompt rev3

Verdict

Prompt rev3 is NOT approved for Agent dispatch yet.

Rev3 fixed the major registry/species field-resolution gap. The semantic registry now covers TAC/IU fields and registry/species fields. This is the correct structural pattern.

However, there are still two issues to patch before dispatch:

1. G10 still contains fuzzy ranking language.
2. Traceability metadata/title still says rev2 in places although the body says rev3.

What rev3 fixed well

1. Registry/species concepts were added to the semantic registry.
2. GATE-0 now includes registry tables.
3. G7 now uses concept IDs for species/composition/governance instead of direct registry field names.
4. composition_level, management_mode, and governance_role are no longer required literal columns.
5. G7 correctly produces grouped evidence and leaves interpretation to GPT/User.
6. TAC/IU field references remain concept-based.
7. Read-only/dry-run boundaries remain intact.

Remaining blockers

1. G10 still has fuzzy ranking language

G10 says:

Rank by: diversity (more section_types = more structural coverage for pilot) and manageable size (not the largest).

This is not deterministic. “manageable size” and “not the largest” invite Agent judgment.

Patch G10 so it outputs candidate metrics only, not a fuzzy ranked recommendation.

Acceptable deterministic options:

• Output all publications with metrics and label all candidate_not_approved; or
• Sort deterministically by declared numeric metrics, e.g. member_count ASC, diversity_count DESC, with no semantic claim that the result is best; or
• Produce buckets such as smallest_by_member_count, highest_diversity, median_member_count_if_computable, all labelled candidate_not_approved.

Do not use “manageable”, “not largest”, “best”, “recommended”, or similar subjective criteria.

2. Traceability mismatch

The content header says prompt rev3, but metadata/title still says rev2 in places. This can confuse future searches/reviews.

Patch title/metadata visible text if possible so the document consistently identifies itself as rev3.

Required patch

Patch prompt to rev4.

Scope:

• G10 only, plus traceability/title/version wording.
• Do not restructure the semantic registry.
• Do not change G1–G9/G11 unless required for wording consistency.
• Do not add new decisions.

Status

phase5_design=ACCEPTED_DIRECTIONALLY
phase5_dryrun_prompt_rev3=NOT_APPROVED_FOR_DISPATCH
reason=G10_fuzzy_ranking_and_traceability_mismatch
agent_dispatch_allowed=false
migration_allowed=false
seed_allowed=false
backfill_allowed=false
next_action=OPUS_PATCH_PHASE5_DRYRUN_PROMPT_REV4_G10_DETERMINISTIC_AND_TRACEABILITY