What problem does it solve? Models that work in training often fail silently in production because of unversioned artifacts, missing promotion gates, untested rollback paths, and monitoring that tracks infrastructure health but not prediction quality. This Skill diagnoses the weakest link in a model lifecycle and scopes the smallest fix that closes the gap. ## Core Features & Use Cases - Lifecycle Gap Analysis: Traces the end-to-end path from training run to production endpoint and identifies nondeterministic builds, unversioned artifacts, or unenforced promotion gates. - Registry and Rollout Design: Designs model registries with lineage from raw data to deployed weights, plus canary, shadow-traffic, and rollback strategies with bounded blast radius. - Drift and Quality Monitoring: Distinguishes infrastructure health (latency, error rate) from model health (prediction distribution shift, feature drift) and wires alerts to each. - Use Case: A team deploys models by copying files to a serving host with no versioning. Use this Skill to design a versioned registry, an enforced promotion gate, and a pre-tested rollback path before the next release. ## Quick Start Ask the agent to audit your model deployment pipeline and design a promotion gate with a tested rollback path.
