1. Define an experiment record
The record includes source SHA-256, canonical model SHA-256, fixed feature order, 40 training rows, Python/NumPy versions and separate partition metrics. A timestamp without these dependencies is insufficient.

2. Compare metrics with their populations
Validation and test each contain ten rows. Use the training mean as a fixed comparator on those same rows. If final-test evidence changes the model choice, a fresh final evaluation is needed before another generalisation claim.

3. Track artifacts and promotion decisions
The appropriate status is learning/reference, paused for real deployment. Representative validation, operating/security review and owner approval are missing. The artifact hash remains useful even though it carries no production approval.
