Opus Post-Dry-Run Review — P3D Pack 1 Phase 4C PASS

Date: 2026-05-11 Author: Opus 4.7 For: GPT-5.5 Thinking / Hội đồng AI — thẩm tra + quyết định species + next steps Run: p3d-pack1-phase4c-dryrun-2026-05-11-opus47 Report: ...reports/p3d-pack1-phase4c-species-mapping-dryrun-report.md

---

1. Verdict: Phase 4C dry-run PASS — evidence complete, 3 decisions pending GPT/User

Agent gathered all 10 goals. GATE-0 all passed. Deterministic labels applied. Scale brittleness self-audit passed. No mutations. Phase 4B snapshots re-verified (all still hold).

---

2. Agent compliance

Check	Result
GATE-0 (6 preflight checks)	✅ All PASS
G1-G10 all attempted	✅ G10 correctly reported FIELD_ABSENT for tac_logical_unit.unit_kind
Zero writes	✅
No species/field from memory	✅ All from live queries
Column introspection per table	✅ 5 tables introspected, FIELD_ABSENT reported where needed
Phase 4B re-verified	✅ 10/10 findings re-verified live
Deterministic labels G8	✅ 8 PLAUSIBLE / 32 CONFLICT / 0 UNKNOWN / 0 N/A
Exact G7 match (no fuzzy)	✅ Subordinate = singleton, exact value
Scale brittleness self-audit	✅ 9/9 passed
All outputs candidate_not_approved	✅

---

3. Key evidence summary

3.1 — 8 PLAUSIBLE candidates, but semantic fit is weak

All 8 PLAUSIBLE species are management_mode=observed (matching IU's governance_role=observed). But semantically:

• governance_infra (atom) — best fit for IU as pilot INFRASTRUCTURE (tables are governance plumbing). But IU's CONTENT is law text, not governance infra. Temporary gom, not permanent home.
• 7 others (ai_support, business_support, checkpoint_support, directus_field, help_center, pivot_result, system_log) — no semantic connection to information units.

SPE-LAW (law, compound, governed) is CONFLICT because it's governed while IU is observed. Yet SPE-LAW is the best SEMANTIC match for law content.

3.2 — Discriminator is dormant system-wide

0/153 species_collection_map rows use discriminators. fn_birth_registry_auto doesn't read them. Any discriminator-based strategy needs either function patch (out of scope) or post-birth correction.

3.3 — law_unit doesn't exist as live data yet

information_unit.unit_kind currently has only design_doc_section (12 rows). law_unit exists only in dot_config vocab. tac_logical_unit has no unit_kind column. First consumer of law_unit would be Phase 5 migration.

3.4 — Composition splits 5 atom / 3 molecule in PLAUSIBLE set

Choosing atom = no containment assertion. Choosing molecule = containment assertion. Current 12 pilot rows have zero nesting evidence (parent_or_container_ref all NULL). TAC has real nesting (depth 2). Phase 5 nesting decision determines correct composition.

---

4. Three decisions for GPT/User

Decision A: Species choice

Option	Species	Composition	Governance match	Semantic fit	Requires
A1: gom governance_infra	governance_infra	atom	✅ observed=observed	⚠️ infra, not law content	Nothing — just INSERT species_collection_map
A2: gom + future promotion	governance_infra now, then QT-005 → SPE-LAW or new species	atom now, compound later	✅ now; ✅ later	Temporary gom, permanent via QT-005	Phase 4C seed now + Phase 5+ QT-005
A3: promote to governed + SPE-LAW	law	compound	✅ governed=governed (after promotion)	✅ law content IS law	QT-005 promotion + species_collection_map
A4: new dedicated species	(new, e.g., information_unit)	(choose)	✅ (choose management_mode)	✅ exact match	New entity_species row + mapping
A5: defer	(none)	(none)	n/a	n/a	None, but Điều 29 violation stays open

Decision B: Phase 5 nesting

Option	Migration approach	Composition consequence
B1: preserve TAC nesting	IU rows get parent_or_container_ref populated	IU with children → NOT atom (molecule or compound)
B2: flatten	IU rows are siblings, nesting as metadata only	atom is fine
B3: decide per section_type	Some sections contain others, some don't	Mixed composition — species-per-unit_kind needed

Decision C: QT-001 scope

Option	What gets backfilled
C1: backfill 12 pilot rows	After species choice (A), update 12 birth_registry rows per QT-001 5-step
C2: defer backfill to Phase 5	Backfill ALL rows (pilot + migrated) in one pass post-migration
C3: backfill pilot + migrate	Backfill 12 now + 86 during migration

---

5. Opus assessment (options, NOT recommendation)

If I had to pick, I'd lean toward A2 (staged): governance_infra gom NOW (Điều 29 compliance, minimal scope, no governance change), then reassess at Phase 5 when nesting decision is made.

But I explicitly do NOT lock this as recommendation per GPT's repeated instruction that design decisions must come from evidence + GPT/User review, not from Opus preference.

Rationale for presenting A2 as a lean (not a lock):

• A1 (gom only) solves Điều 29 but leaves no future path documented
• A3 (promote + SPE-LAW) is the cleanest long-term but requires governance promotion before migration — adds scope
• A4 (new species) is accurate but adds entity_species row before knowing composition
• A5 (defer) leaves Điều 29 violation open

A2 documents the future path (discriminator rows + QT-005 plan) while solving the immediate compliance gap with minimal scope.

But GPT may see a different balance. The evidence is complete — the decision is theirs.

---

6. Handoff checkpoint

This is a natural handoff point. Context summary for next session (if needed):

What's done: Phase 4 vocab (14 keys committed, planner law_unit=plan_ok). Phase 4B discovery (40 species, 153 mappings, fn_birth_registry_auto documented). Phase 4C dry-run (8 PLAUSIBLE candidates, discriminator dormant, law_unit not live yet).

What's pending: GPT/User choose species (A1-A5) + nesting (B1-B3) + backfill scope (C1-C3). Then Phase 4C executable prompt (seed species_collection_map + QT-001 backfill).

What's blocked until decisions: Phase 5 TAC→IU migration.

phase4c_dryrun=PASS
evidence_complete=true
agent_compliance=FULL
species_decision_locked=false
composition_decision_locked=false
decisions_pending=A_species_B_nesting_C_backfill
next_action=GPT_USER_DECIDE_THEN_PHASE4C_EXECUTABLE_OR_HANDOFF