Model Serving & Monitoring Audit

Operation

Inspect the named boundary and report supported findings. Do not edit product code. Include concrete evidence, impact, the owning source, one remediation direction and a meaningful validation route. Severity follows actual impact, not a category example.

Goal and scope

Audit model serving across artifact identity, preprocessing, routing, batching, concurrency, fallback, observability, and safe rollout.

Domain invariants

• Serving loads the intended model/preprocessing/schema/threshold versions atomically and reports their identity.
• Admission, batching, pools, concurrency, timeouts, cancellation, retry, and memory/GPU limits are bounded.
• Fallback or shadow routing preserves semantics, tenant/privacy controls, and truthful status; it never silently changes product policy.
• Latency, errors, saturation, input/output quality, drift, and cost are observable with privacy-safe dimensions and rollback controls.

Audit method

1. Trace request routing through preprocessing, model selection, batch/queue, inference, postprocessing, persistence, and response.
2. Check artifact/config skew, cold start, overload, partial batch, timeout, cancellation, provider outage, fallback, and mixed-version rollout.
3. Inspect health/readiness semantics, autoscaling signals, cache identity, canary/shadow evaluation, alerts, and operator disable.
4. Separate performance evidence from quality/drift evidence.

Priority model

• P0: unsafe action, privacy or tenant leak, materially wrong irreversible decision, corrupted model/data lineage, or critical service failure.
• P1: a reachable quality, grounding, evaluation, serving, cost, or governance defect with clear product impact.
• P2: a lower-risk but concrete robustness, observability, dataset, or maintainability issue.