GPT Review — After Registries-Pivot P0+P1 Rehearsal and Next Macro

Date: 2026-05-31 Reviewer: GPT Council

Current status

The P0+P1 rehearsal passed and proved the corrected view layer is technically feasible under BEGIN..ROLLBACK, but production implementation is still gated.

Key findings

• Count integrity is still FAILED, but now leaf-scoped and quantified: net_gap +36,783 / 3 drift rows.
• Unscoped counting was wrong because rollup/meta rows double-count.
• Corrected scalar-EXISTS view strategy compiled and reproduced the leaf invariant.
• Dynamic drilldown mechanism is sound but data wiring is blocked because pivot_definitions.parent_code is NULL on all 37 rows.
• Label/grouping substrate exists and must be reused.
• Threshold registry gap exists: max ungrouped threshold must be data, not frontend constant.
• Pin/ghim registry is absent; registry_pin is NEW but approval-gated.
• Phantom remains LAW_DEFINITION_GAP and must be source_model-aware.
• Live /knowledge/registries and /api/registry/health still violate Đ28 and must be retired/replaced after the pivot-backed replacement is ready.

Direction

The next macro should be larger than a narrow commit. It should combine:

1. human-ratification packet;
2. count-integrity commit-ready SQL pack;
3. missing pivot bundle design/rehearsal;
4. drilldown hierarchy wiring design/rehearsal;
5. threshold/label and pin registry design/rehearsal;
6. phantom policy draft;
7. Directus/API exposure readiness;
8. static UI preview plan;
9. legacy retirement plan;
10. implementation prompts.

No production commit should happen unless a valid approval artifact is found. If no approval is found, the agent must produce a commit-ready pack and rehearse only.

Guardrails

• No hardcode or disguised hardcode.
• No frontend count logic.
• No production Nuxt change.
• No PG COMMIT without approval.
• BEGIN..ROLLBACK allowed for proof.
• Every count pivot-backed or PIVOT_MISSING.