Set when the shipped finalist's family is hidden for its holdout calibration (a probability task whose slope is
outside 0.8–1.2, or could not be measured): the model stays Draft and the promotion gate never moves it to
Validated or Published, with or without a sign-off. Null when nothing is held (a kept baseline member is not
judged by the finalist's calibration report).
The driver key of the shipped artifact. It can differ from the driver of the model row's AlgorithmID: a shipped
count or non-negative GLM (poisson_glm, tweedie_glm) or forecast baseline (naive, seasonal-naive) has no
MJ: ML Algorithms row, so the row names the pipeline's algorithm and this key holds the truth.
What the gate shipped as this model: the finalist's name, or the simple baseline member kept (glm, oneFeature, naive, seasonalNaive).
The run's leakage audit: the top feature of the finalist's held-out importance and its share, whatever shipped.
A kept simple baseline stores no importance of its own (FeatureImportance is {}), so the promotion gate reads
this audit to judge it against the pipeline's current dominance threshold. Null when the trainer measured no
importance (a forecast reports none).
What the Lineage of a model trained under an Evaluation Protocol adds: the protocol it was judged under and the format of its artifact (the trainer's envelope, which carries its own in-fold encoding).