OptionalExpectOutputs whose correct answer is an abstention: withheld below the confidence floor, no answer, or an Expected label.
Per output, the accepted labels (first = primary); for an ExpectAbstain output, accepted abstain values.
Stable case ID, e.g. "g01".
The record fields the pipeline reads. unknown because the values are whatever column types the pipeline's
entity has; the scorer never reads them, it only passes them to the pipeline and hashes them.
Where the labels came from: human (an app user), model-consensus (two model labelers agreed),
model-adjudicated (a third labeler decided), specification (fixed by the pipeline's written rule, not labelled).
OptionalRequiredA must-pass case: when any output it scores is wrong, the gate value is 0 whatever the accuracy.
One golden case (stored in the test's InputDefinition): the record input and, per output, the accepted labels.