StaticCalibrationWhether a family is hidden for its holdout calibration, so it is never offered. Only the probability tasks the trainer calibrates (binary and multiclass, TaskTypeRules.IsClassificationTask) can be hidden: a slope outside 0.8–1.2, or one that is not a finite number, hides the family, and so does a null report, because the sidecar returns none when the slope could not be measured on the holdout rows (constant predictions, separated classes, a fit that does not converge) or when the protocol declares no calibration. Any other task is never hidden: the trainer calibrates none of them.
the completed result's holdout calibration report
the task the run trained
StaticCodeWhat a CodeVersion lacks for lineage: missing library versions and a thread count below 1. Empty = complete.
StaticFeatureA trainer run's held-out permutation importance as the stored FeatureImportance map: each source column's share
of the total positive loss increase. It is the same quantity, in the same format, the sidecar's legacy /train
response stores (_extract_importance in the sidecar's main.py): a decrease counts as zero, and every share is 0
when no column has a positive increase. An empty list (a forecast reports none) is {}.
StaticHoldoutA completed run's holdout scores by metric: the primary score and every secondary one.
StaticImportanceWhich kind of number FeatureImportanceOf holds: held-out permutation shares, or none when nothing was measured.
StaticParseCheck a sidecar's POST /trainer answer against the contract: the typed result, or one problem per mismatch
(<path>: <message>). Unknown keys are refused, as the sidecar's own models refuse them.
StaticToThe result with every returned artifact replaced by the MJ: Files row it was stored as (fileIds maps candidate
to file ID), so the stored result carries no artifact bytes. Generic over the result shape, so a
TrainerCompleted becomes a StoredTrainerResult.
Rules over the trainer contract.