Whether the user may ACT on this prediction (the gate). false for Poor or unmeasured models.
Guidance for how much to rely on it (grade-specific).
When canAct is false, the plain reason actions are locked; otherwise null.
The plain reliability grade.
The metric that drove the grade, for the "For analysts ▸" disclosure. null if none usable.
Set when a classifier is no better than always guessing the most common class (or predicts one class): the rule
in force, whether it capped the grade, and the numbers. Null otherwise. It is computed and carried here but not yet
rendered or recorded anywhere: under warn, where it does not cap the grade, nothing tells the user.
Headline in plain words, e.g. "Right about 8 out of 10 times."
0–1 normalized trust score, for a meter/visual (NOT a probability).
True when no usable quality metric was found — trust is unknown and actions are gated off.
The stored reason the verdict is withheld (e.g. "not evaluated as a forecast"); null otherwise.
The full plain-language verdict for a model, derived from its metrics.
The mixed member casing is deliberate: the camelCase members predate the PascalCase convention and are kept so existing consumers keep compiling; members added later (
WithheldReason,MajorityBaseline) follow the convention. Do not rename either half to match the other.