Lab data does not represent every production environment
Lighting, camera angle, distance, background and equipment wear can change the input distribution. Evaluation sets must cover the real scenarios of each operating location.
A decision threshold belongs to the workflow
Aggregate accuracy is not enough. The cost of false positives and negatives, human review capacity and response time determine how model output should enter operations.
Measurement continues after deployment
Image quality, distribution drift, human escalation rate and final outcomes should be recorded so degradation is visible before it becomes an operational incident.