Architecture and MVP Boundary
What is real
- Authenticated ingestion endpoints for metrics and deployment events.
- PostgreSQL persistence with idempotent and revision-safe ingestion behavior.
- Durable jobs table with worker claiming, retries, and redacted failure state.
- Deterministic tracking_integrity_failure@1 evaluation and suppression logic.
- Deterministic deployment correlation scoring and exposure calculation.
- Incident lifecycle with append-only incident event history.
- Asynchronous replay runs with persisted run-state polling and session-aware contention handling.
Processing paths
- Normal ingestion path: source payloads enqueue durable evaluation jobs and are processed by the worker process.
- Contest replay path: replay orchestration persists demo records and drains only demo jobs inline to keep deterministic timing.
- Both paths execute the same persisted rule evaluation, incident creation, correlation, and recovery logic.
AI investigation boundary
- Deterministic rules create incidents, calculate exposure, score deployment correlation, and verify recovery.
- AI is limited to an on-demand investigation brief generated from persisted incident evidence.
- AI may summarize competing interpretations and inspection steps, but it cannot claim causation or control spend.
- If model configuration is unavailable or output validation fails, CatchDrift falls back to deterministic guidance.
What is replay-controlled
- Campaign source data values are controlled for deterministic contest replay.
- Provider OAuth connectors for ad and affiliate platforms are not included.
- Automatic remediation is intentionally out of scope.