01 — Overview & Control

Mission: LAW_CAPABILITY_DISCOVERY_SECOND_PASS_CROSS_CHECK_AND_DECISION_GATE Date: 2026-05-30 · Class: second-pass anti-forgetting verification + decision gate Execution mode: READ-ONLY. Zero PG / Directus / Qdrant mutation. KB docs are the only artifacts written. Host/DB: contabo VPS · live PG read via query_pg (AST-validated read-only role, READ ONLY tx, 5s timeout, LIMIT 500). Substrate lives in directus.public (274 tables) + iu_core.

Why this pass exists

The first pass (law-capability-discovery-official-lists-automation-2026-05-30, 13 docs) produced a strong reuse-first inventory. The GPT Council accepted it but explicitly forbade starting implementation until a second pass independently verifies or falsifies it and proves no existing law / capability / "musician" was missed. This is a conductor problem, not a new orchestra problem: identify every existing instrument before composing.

What this pass did

1. Read all 13 first-pass docs + the GPT review in full.
2. Independently re-ran the headline claims against live PG — not trusting the first pass's numbers, re-querying each.
3. Cross-checked laws from 4 angles: KB semantic search, normative_registry, law_catalog, and prior reports/handoffs (incl. S164/S165/S165-KB/S166-KB/S178/Đ37-38 bootstrap that the first pass did not cite).
4. Verified pivot, DOT, birth, IU/KG, event, approval, and UI-Preview findings with live objects.
5. Built anti-forgetting ledger v2 and a decision gate classifying every proposed next action.

Headline verdict

PASS — first pass INDEPENDENTLY VERIFIED, with corrections. ~40 discrete facts re-confirmed exact against live PG. The reuse-first thesis stands: the pivot+counting engine, birth gate, meta_catalog, approval spine, event substrate, IU/KG substrate and DOT governance all exist live and must be reused, not rebuilt. No new law beyond the 15 mandatory was found; no capability that overturns the plan.

But the second pass surfaced 5 corrections / sharper findings the first pass got wrong or under-stated (detail in docs 02/04/05/06/10):

1. trg_auto_sync_registry_counts is ENABLED (tgenabled='O'), NOT "disabled." The first pass carried a stale KB (s146) claim. The count-integrity defect (cross_check LỆCH) is real but mis-attributed — re-enabling a trigger is the wrong fix.
2. CAT-DOT=307 vs CAT-006=309 — the two meta_catalog rows counting dot_tools disagree by 2 (and CAT-006 is internally LỆCH 309 vs count_b 163). Sharper than first pass's single-figure note.
3. birth_registry live = 959,372 (first pass said 954,161) and event_outbox = 166,422 (first pass 165,403) — hand-pinned counts on high-churn tables are already stale. Reinforces: never hand-count, always pivot.
4. CAT-007 'page' = 37 vs count_b 52, LỆCH — a second UI dual-count beyond the manifest-vs-ui_pages gap.
5. First pass omitted the foundational report family S164/S165/S165-KB/S166-KB/S178/Đ37-38-bootstrap. Mostly corroborating, but S165-KB's "reference-table-replaces-CHECK + config-table-replaces-hardcode" pattern and S178 (DOT Repair Governance, origin of Đ35 v5.2 paired-test/fix_repair_dot) deserve explicit anti-forgetting capture.

Package contents (13 docs)

1. Overview & control (this doc) · 2. First-pass verification summary · 3. Law cross-check matrix · 4. normative/law_catalog drift analysis · 5. Pivot/counting cross-check · 6. DOT governance cross-check · 7. Birth/collection cross-check · 8. IU/KG/profile cross-check · 9. UI Preview official-list cross-check · 10. Anti-forgetting ledger v2 · 11. Decision gate classification · 12. Recommended next macro · 13. Self-review.

Forbidden-action compliance

No INSERT/UPDATE/DELETE/DDL. No volatile function executed (pivot/refresh/birth/test functions inspected by definition only or counted via plain SELECT). No Directus/Qdrant write. No write_file/ingest_document. UI accessed via first-pass HTTP record + KB. Only KB upload_document for this report package. Entry counts == exit counts (no write path).

Builds on [[law-capability-discovery-official-lists-automation-2026-05-30]] (first pass) and the GPT review.