What problem does it solve? ML work often suffers from metric shopping, data leakage, untracked runs, and notebook-only results that cannot be reproduced or handed off. This Skill structures machine learning work as a scientific workflow with locked metrics, tracked experiments, and validated candidates. ## Core Features & Use Cases - Locked Metrics and Baselines: Defines primary offline metrics, online decision metrics, guardrails, and a simple baseline before any experiment results are seen. - Bounded Hypothesis Loop: Generates hypotheses with refutation criteria, trains and validates iteratively, and stops on metric success, compute caps, or diminishing returns. - Reproducible Handoff: Packages model artifacts, data lineage, tried/rejected ledgers, limitations, and deployment readiness for research or production decisions. - Use Case: A data scientist testing a new ranking model locks the offline metric and guardrails, runs tracked experiments against a baseline, validates the best candidate for leakage and fairness, and hands off a deployment-ready package. ## Quick Start Use the ml-experiments skill to frame my churn prediction problem, lock a primary metric, and run a tracked baseline experiment with a reproducible handoff.
