# PRE-AI LOCAL COMPLETION V3

STATUS=COMPLETE_OFFLINE_WITH_EXTERNAL_BLOCKERS

## Measured Product State

DATABASES=5
PRODUCTS=35066
FISCAL_CELLS=55756
DETERMINISTIC_GATE_CLOSED=true
FIREBIRD_AI_WRITES=0
AI_WRITE_ALLOWED=false
EXTERNAL_API_CALLS=0

KB_PRODUCTS_BEFORE=30801
KB_PRODUCTS_AFTER=30801
CANONICAL_PRODUCTS=35065 observation identities
KB_CONFLICTS_BEFORE=163
KB_CONFLICTS_AFTER=90 observable audited; 73 historical conflicts lack detail in the seed payload

WORK_ITEMS_BEFORE=750
WORK_ITEMS_AFTER=750
LLM_ITEMS_BEFORE=511
LLM_ITEMS_AFTER=511

## Retrieval

GOLD_RECALL_AT_1_BEFORE=0.7846
GOLD_RECALL_AT_1_AFTER=0.7846
GOLD_RECALL_AT_20_BEFORE=0.9538
GOLD_RECALL_AT_20_AFTER=0.9538
GOLD_MRR_BEFORE=0.8392
GOLD_MRR_AFTER=0.8392
RETRIEVAL_V3=NOT_ACTIVATED_PENDING_FULL_HOLDOUT_REPLAY
LOCAL_EMBEDDINGS_USED=false
LOCAL_EMBEDDINGS_IMPROVEMENT=NOT_TESTED; no model download/GPU introduced

The V3 evidence layer is available but preserves V2 ordering unless a complete
untouched holdout proves improvement. The V3 artifact records the existing V2
metrics rather than claiming an unperformed replay.

## Pipeline Verification

FAKE_PROVIDER_COMPLETE=true
BATCHING_TESTED=5,10,20
RETRY_TESTED=TIMEOUT,429,500
CACHE_TESTED=true; all key dimensions and zero-call hit policy
BUDGET_GUARD_TESTED=true; pre-call and actual usage guard
E2E_SIMULATION=GOLD -> gate -> SILVER -> NOISY -> residual plan
FINAL_E2E_AFTER_PHASE_2=PASS_PLUMBING
FAKE_RESULTS_AUTHORITATIVE=false

Artifacts:

- `docs/KB_CONFLICT_AUDIT_V1.json`
- `docs/NCM_RETRIEVAL_EVALUATION_V3.json`
- `docs/AI_RESIDUAL_PLAN_V3.json`
- `docs/runs/offline_ai/final/OFFLINE_AI_SIMULATION_V1.json`
- `docs/LOCAL_PRODUCT_HEALTH_V1.json`
- `docs/LOCAL_VALIDATION_V1.json`

## Human Operations

COMPANY_QUESTION_CLI=true
GROUP_QUESTION_CLI=true
ENTRY_SCOPE=INTRA|INTER|MIXED_OR_DYNAMIC; MIXED_OR_DYNAMIC does not invent static CFOP_ENTRADA
STORE_IPI_PROFILE=POPULATE_REFERENCE_RATE|LEAVE_NOT_APPLICABLE; rollback is preview-only until operator confirmation
ROLLBACK_TOOL_READY=true; current == expected_after guard emits CONFLICT
FULL_LOCAL_VALIDATION_COMMAND=`python scripts/run_local_validation.py`
HEALTH_REPORT=`docs/LOCAL_PRODUCT_HEALTH_V1.json`

NCM_AI_READY=244 deduplicated work items
CEST_AI_READY=353 work items
EX_TIPI_AI_READY=112 work items
PRODUCT_ROLE_AI_READY=41 work items

AI_PAYLOAD_TOKENS_BEFORE=NOT_MEASURED_FOR_511_ITEMS; final per-item prompt payloads are not present in the local artifact set
AI_PAYLOAD_TOKENS_AFTER=NOT_MEASURED_FOR_511_ITEMS
AI_PAYLOAD_REDUCTION=NOT_MEASURED

## Verification

FULL_TESTS=518 passed, 1 skipped, 1 existing openpyxl warning
SECRET_SCAN=NO_TRACKED_SECRET_PATTERNS
DIFF_CHECK=PASS

## Remaining External Blockers

- Z.ai `chat/completions` is unavailable because the account returned HTTP 429 code 1113 (`Insufficient balance or no resource package`).
- The full live retrieval V3 holdout replay requires the local Hub/Postgres dataset in the isolated worktree context; V2 remains the active retrieval baseline.
- 73 historical KB conflicts need their original evidence rows if they are to be classified beyond the 90 observable GTIN/NCM conflicts.
- Human company/group decisions remain intentionally unresolved.

VERDICT=LOCAL_PIPELINE_OPERATIONALLY_READY; REAL_LLM_EXECUTION_AND_SEMANTIC_WRITES_REMAIN_BLOCKED_BY_POLICY_OR_EXTERNAL_QUOTA

## PRE-ZAI Closure (2026-09-21)

`docs/PRE_ZAI_FINAL_STATE_V1.json` is the current final-state artifact.
Historical conflict reconstruction inspected the available seed payload fields
(`gtin`, fingerprint, normalized description, unit, NCM, and observations), but
the original 73 conflict identities are absent. All 73 therefore remain
`INSUFFICIENT_EVIDENCE` and `PRESERVE_CONFLICT`; the 90 observable current NCM
conflicts were not changed. The COMPANY session is review-only, with 19 groups
(14 existing COMPANY questions plus 5 explicit per-database IPI profile
choices), and reports real local IPI impact: CONSTRUCAO 0, MMMODAS 1, RINCAO 0,
TESTESOFICIAIS 198, UNIAO 1 currently filled values eligible for a future
operator-approved rollback plan. No context answer was persisted.