2.1 KiB
Autonomous Operation Policy V1
Operational Contract
Each product ends in exactly one state:
COMPLETECORRECTEDBLOCKED_MISSING_EVIDENCEBLOCKED_CONFLICTBLOCKED_COMPANY_CONTEXTBLOCKED_SCHEMAFAILED_TECHNICAL
COMPLETE and CORRECTED products continue normally. Blocked products never
receive uncertain writes. FAILED_TECHNICAL is retried by the job system and
does not stop processing of other products.
Partial Processing
The Agent processes every product independently and reports per-product state
counts. Safe blocked products do not stop processing, heartbeat, new jobs, or
reconnection. AGENT_HEALTH=HEALTHY is valid when SAFE_BLOCKED > 0 and
FAILED_TECHNICAL=0.
Go-Live Gate
GO_LIVE_ALLOWED=true only when all are true:
- preflight passes;
- company context is resolved;
- mandatory infrastructure is healthy;
- no unsafe write path is available;
- blockers are explicitly registered as safe blocked states;
- technical failures are below the configured threshold.
One hundred percent product completeness is not required to start an Agent. This policy authorizes no deployment by itself.
Blocker Registry And Retry
The Hub stores one blocker per stable fingerprint with tenant/company/product,
field, evidence, timestamps, retry count, next retry, and resolution source.
Identical input fingerprints are BLOCK_CACHE_HIT and do not create duplicate
rows, notifications, AI work items, or external calls. Reprocessing occurs when
product/context/dataset/KB/engine input changes or a retry schedule is due.
The exception API supports listing/filtering, CSV export, reprocess job creation, and explicit resolution. A successful later evaluation automatically closes all unresolved blockers for that product.
Current Policy Facts
AI_REQUIRED_FOR_MANDATORY_COMPLETENESS=0
EXTERNAL_PRODUCT_PROVIDER_REQUIRED_FOR_MANDATORY_COMPLETENESS=0
AI_WRITE_ALLOWED=false
FIREBIRD_AI_WRITES=0
The 511 AI work items are optional enrichment. No provider purchase, AI call, or fiscal inference is needed for operational start under this policy.