Brief
Log 3+ runs (params/metrics/artifacts) in the tracking harness, compare them, and submit a run-comparison report plus a model card for the promoted run.
Dataset
Churn-model training split from the starter (train/holdout hash recorded per the m0 checklist). Same model family served later by the flagship.
Starter guardrails (plan 88)
Use the tracking harness only; log params, metrics, artifacts, seeds, and data checksum per run; no registry writes outside the harness.
100-pt rubric
Run completeness (3+ runs, all fields) 30; comparison reasoning 25; model card (intended use, limits, metrics) 25; reproducibility fields (seeds + checksum) 20.
Deterministic checks -> LLM review
Checks: >= 3 runs logged; promoted run has a model card; checksums present. LLM review grades the comparison reasoning and the card's limits section.