Module 12: deployment and monitoring

{
  "module": 12,
  "real_heldout_rows": 243,
  "batches": 4,
  "model": "logistic",
  "threshold": 0.3,
  "replay_order": "source record order, not a measured event-time stream",
  "labels": "Offline known labels for delayed-outcome monitoring rehearsal",
  "monitoring": "Measured duration, input statistics, and outcome metrics; no fabricated telemetry"
}

Batches contain actual held-out records. Dataset order is not a temporal sequence; batch differences are not evidence of time drift.

Measured local durations depend on hardware/warm-up and are not production latency guarantees.

A mean-shift diagnostic prompts investigation; it does not prove concept drift. Operational use needs privacy, ownership, delayed labels, and rollback.

The Streamlit app uses this checked inference function and a local-LLM form that rejects invalid output schemas.