docs: record pre-zai closure design

This commit is contained in:
OpenCode
2026-09-21 16:50:45 -03:00
parent b10131953c
commit a6492f936e
2 changed files with 105 additions and 0 deletions
@@ -0,0 +1,37 @@
# PRE-ZAI Blocker Closure Design
## Goal
Close every locally verifiable blocker before the real Z.ai GOLD benchmark,
without changing the 90 valid NCM conflicts, inferring human facts, calling
external AI, deploying, or enabling Firebird AI writes.
## Design
The historical conflict audit joins every available seed-payload observation by
GTIN and fingerprint, preserves descriptions/tenant/product/NCM evidence, and
classifies each historical row as `RESOLVED_EQUIVALENT_PRODUCT`, `FALSE_MATCH`,
`VALID_NCM_CONFLICT`, `STALE_OBSERVATION`, or `INSUFFICIENT_EVIDENCE`. Missing
evidence remains unresolved and never selects an NCM by majority.
The operator session exposes only COMPANY questions first. It is guided,
cancelable, review-first, and produces a context import artifact only after
explicit confirmation. IPI options include real local impact counts where the
source artifacts expose them; rollback remains a plan until the operator
chooses it. `MIXED_OR_DYNAMIC` produces no universal CFOP.
Final health derives from existing artifacts and the regenerated residual plan.
The report uses `COMPANY_CONTEXT_COMPLETE=false` until answers are actually
persisted, and labels GROUP/PRODUCT decisions as human-pending rather than
pipeline failures. Retrieval V2 remains active unless a complete holdout proves
V3 improvement.
## Safety and Acceptance
- Current 90 conflicts remain `NCM_CONFLICT` unless objective evidence changes a
specific row.
- Historical rows without evidence are `INSUFFICIENT_EVIDENCE`.
- All reports state `FIREBIRD_AI_WRITES=0` and `AI_WRITE_ALLOWED=false`.
- No answer is auto-persisted by the guided session.
- The final artifact contains no placeholder TODOs and lists only external AI
quota/credential plus deliberate human decisions as remaining blockers.