ML Model Development

Dataset identity/version, split and leakage rules, canonical evaluation protocol and its repository reference. Keep important caveats here; link detailed protocols instead of copying them. Describe verification as reported or independently checked, with who checked what and when.

Current comparison reference or selected model, why it was selected, evaluation setting and run/artifact pointer. Explicitly say when no model has been evaluated or selected. The latest run is not automatically best. Development results are not independent test results.

Lead with one short sentence about the active work. Preserve useful detail after it. Move completed work into learnings rather than presenting it as active. Check factual accuracy and relevance when the project changes direction.

Lead with the next useful action. List blockers and unresolved questions only when known. Remove completed actions and carry adopted prerequisites forward. Never confirm this section just to clear a warning.

Latency, compute, data access, deployment, licensing and other practical limits. Record required infrastructure as unmet until supported by evidence; an experiment tracker is optional unless the project explicitly requires one.

Repository/commit, dataset version, protocol, run, model or supporting document pointers. Label planned, reported existing, checked or unavailable; record who checked and when. Local paths may not be accessible to another assistant. Do not invent identifiers, copy metrics tables, or store credentials.