GPT Review — File 13 Description Policy Tiering + Agent Investigation Prompt

Date: 2026-05-04 Reviewer: GPT-5.5 Thinking / Incomex Hội đồng AI Reviewed: knowledge/dev/laws/dieu44-trien-khai/design/13-iu0-description-policy-tiering-and-pack2b-f6-preflight.md rev 1

Verdict

PASS DIRECTION. APPROVE READ-ONLY AGENT INVESTIGATION AFTER PROMPT PATCH.

Opus correctly identified the foundational issue: Description Governance cannot be one-size-fits-all. Tiered Description Policy is necessary before Pack 2B creates IU rows.

No runtime, DDL, law patch, or Pack 2B opening was performed. Scope discipline is good.

Law / Constitution check

No blocker conflict in direction.

• Đ3: Current law already distinguishes architecture entities vs operational records at enrichment scope. Tiering formalizes this instead of contradicting it.
• Đ4: Birth guard currently enforces description mostly by governance_role; tiering proposes a more precise policy while preserving strict enforcement for Tier A.
• Đ43/H11: H11a/H11b need machine-checkable policy to avoid false positives for IU rows.
• Điều 20 / NT15: design-before-runtime respected.
• Đ44: Pack 2B remains blocked until F6/description policy/birth path are designed.
• HP NT9: unknowns go to read-only investigation, not assumption.

What Opus did well

1. Correctly defined Tier A/B/C.
2. Correctly separated detailed description enrichment from structured metadata / graph semantics.
3. Correctly identified that IU rows should not require agent-written detailed description by default.
4. Correctly identified that H11a, H11b, birth guard, Đ3, Đ4, Đ43, and Pack 2B are all affected.
5. Correctly did not patch laws or run runtime.

Issues to fix before dispatching Agent

P1 — Q7 SQL precedence bug

Current Q7 query:

WHERE table_name IN ('information_unit', 'unit_version')
  AND column_name LIKE '%profile%' OR column_name LIKE '%description%';

Because AND binds before OR, this may return any table with description-like columns. Must use parentheses:

WHERE table_name IN ('information_unit', 'unit_version')
  AND (column_name LIKE '%profile%' OR column_name LIKE '%description%');

P2 — Use pg_get_functiondef, not only prosrc

For Q6, request:

SELECT pg_get_functiondef('fn_description_birth_guard'::regproc);

This is safer and includes signature/schema.

P3 — Check actual H11 registration/executor source

Agent should inspect where H11a/H11b live in runtime:

• system_health_checks rows for H11a/H11b, if present.
• query files under knowledge/current-state/queries/* are docs, but runtime executor may use DB rows or scripts.

Add read-only query:

SELECT code, name, severity, query_sql, is_active
FROM system_health_checks
WHERE code ILIKE '%H11%' OR name ILIKE '%Description%';

If columns differ, inspect schema first.

P4 — Check Directus/schema exposure implications before recommending collection_registry DDL

Option 1 may be right, but it is DDL. Agent must inspect whether collection_registry has flexible metadata fields and whether Directus fields already include extension columns.

Add:

SELECT * FROM directus_fields WHERE collection = 'collection_registry' ORDER BY sort, field;

And inspect JSONB/config columns.

P5 — Check existing label taxonomy/facets more safely

Q3 should query both taxonomy and taxonomy_facets schemas first, because column names may differ. Avoid assuming code, name exist without schema check.

P6 — Clarify Tier B is not “no metadata”

Tier B is exempt from detailed/free-text description, not exempt from metadata. Agent/report must say IU rows still need structured metadata and graph edges.

P7 — Option 1 should be called “candidate”, not recommendation final

File 13 says pre-investigation recommendation = Option 1. This is acceptable, but Agent prompt/report must not treat it as predetermined. Investigation should compare field vs label vs config vs mapping table, including the possibility of using existing h11b_exclude_species short-term.

P8 — Add law/document read list in Agent prompt

Agent should read at least:

• file 13 rev1;
• description-enrichment-guide.md;
• H11a/H11b query docs;
• Đ3/Đ4/Đ43 relevant excerpts if accessible.

Directive

Ask Opus to patch file 13 to rev2 with the prompt corrections above, then prepare a final read-only Agent prompt. Do not dispatch until User/GPT approves the exact rev2 prompt.

Do not open Pack 2B. No DDL. No law patch. No runtime write.

Next expected flow

1. Opus patches file 13 rev2.
2. GPT/User reviews the exact Agent prompt.
3. If approved, Agent runs read-only investigation.
4. GPT reviews Agent report.
5. Opus then updates design/recommendation and proposes law/tool changes.