Data quality
Capture failures are caught, classified, and routed · design goal: nothing silently dropped
How errors are caught
Five layers between the water and the traceability register
| Layer | Catches |
| 1 · Schema validation | Malformed or incomplete payloads rejected at the API, never enter the register |
| 2 · Identifier check | Events without a readable QR/RFID batch ID quarantined to the reconciliation queue |
| 3 · Confidence threshold | CV detections under 90% confidence routed to human review |
| 4 · Heartbeat + watchdog | Silent cameras or sensors flagged within minutes; gaps marked on charts |
| 5 · Offline buffer | Connectivity loss = events held at the edge with original timestamps, replayed on reconnect |
Why this matters for FSMA 204
Error handling is a traceability requirement, not a nice-to-have
An FDA records request is answered from the traceability register. A silently dropped event is a hole in that record. Catching every failure, and proving each one was either recovered or consciously resolved, is what makes the register defensible.
The error queue doubles as an audit trail: every anomaly, when it was caught, how it was resolved, and by whom.