sw-model-registry_skill

Overview

This skill provides a centralized model registry for managing ML model lifecycle: versioning, staging, metadata, lineage, and safe rollback. It integrates with SpecWeave increments so models are automatically registered, documented, and traceable from experiment to production. The registry supports semantic versioning, promotion workflows, and audit-ready metadata.

How this skill works

The registry stores model artifacts, metadata, signatures, environment files, and feature schemas per increment. It exposes operations to register versions, auto-increment semantic versions, promote models across dev→staging→production, and execute one-command rollbacks. Lineage and comparison tools let you trace data→features→experiment→model and compare metrics, latency, and feature changes.

When to use it

• You need a single source of truth for all models and versions
• You want reproducible deployments tied to SpecWeave increments
• You require safe promotion workflows and approval gates before production
• You need audit trails and lineage for compliance or debugging
• You want automatic rollback triggers on metric degradation

Best practices

• Follow semantic versioning: major/minor/patch for breaking changes, features, and bugfixes
• Record full metadata: training data version, hyperparameters, feature list, environment, and increment id
• Use staging and A/B testing (e.g., 10% traffic) before production promotion
• Define approval workflows and approvers for production promotions
• Set monitoring with auto-rollback triggers for error, latency, or accuracy thresholds

Example use cases

• Auto-register models at increment completion so every experiment is reproducible and discoverable
• Run an A/B test in staging then promote the winning model to production with an approval step
• Rollback production to the previous stable version automatically when error_rate or latency crosses thresholds
• Compare two model versions side-by-side to validate metric and feature changes before promotion
• Generate audit reports listing who approved deployments, model versions deployed, and data sources used