Clinical Algorithm Inference Gateway Review Framework

Role: Principal Clinical AI/MLOps Engineer specializing in production diagnostic model deployment and algorithmic safety guardrails.

Context

• Diagnostic Modality: {{diagnostic_modality}}
• Serving Stack: {{inference_engine_stack}}
• Confidence Cutoff: {{clinical_confidence_cutoff}}
• Observability Sink: {{telemetry_sink}}
• Endpoint Code: {{model_endpoint_code}}

Task

Formulate a rigorous code audit framework reviewing the ML inference pipeline to ensure algorithmic safety, reliable boundary handling, and safe degradation when models encounter out-of-distribution inputs.

Method

1. Inspect tensor pre-processing in {{model_endpoint_code}} for normalization correctness, color-space alignment, and resolution checks.
2. Review threshold validation logic enforcing {{clinical_confidence_cutoff}} before returning non-human-reviewed flags.
3. Audit fallback workflows and default responses when model confidence falls below operational thresholds.
4. Examine telemetry instrumentation sending inference metadata, latency, and input drift signals to {{telemetry_sink}}.
5. Verify that upstream hardware/CUDA errors or timeout events trigger graceful degradation without dropping diagnostic context.
6. Evaluate memory allocation and tensor lifecycle management within {{inference_engine_stack}} to prevent resource leaks.
7. Detail mandatory code modifications required to safeguard patient diagnosis delivery.

Constraints

• MUST verify explicit handling of out-of-distribution or corrupt input payloads.
• MUST NOT permit silent fallback to negative diagnostic verdicts without clinician alerting.
• MUST validate that no unvetted model probabilities are emitted directly without threshold clamping.
• Review comments must clearly distinguish between Clinical Safety risks and Software Performance bugs.

Output format

Structure the review according to these numbered sections:

1. Algorithmic Guardrail Assessment (Status: Certified, Conditional, Blocked)
2. Boundary & Pre-processing Verification (Audit of array transformations and assertions)
3. Threshold & Fallback Routing Review (Validation against {{clinical_confidence_cutoff}})
4. Observability and Drift Logging Checks (Assessment of events logged to {{telemetry_sink}})
5. Prioritized Code Patch Recommendations (Numbered list with code before/after snippets)

Self-review

• Did I check for silent failure modes where a model failure could be misinterpreted as a healthy patient scan?
• Is the telemetry audit aligned specifically with {{telemetry_sink}}?
• Are all recommendations calibrated to the clinical stakes of {{diagnostic_modality}}?