Evaluation data controls
Check sentinel keys, declared-link components, controlled perturbations and conditional prevalence assumptions.
Detect accidental answer-key changes with a sentinel
A dataset update changes expected answers without an authorised rubric change.
Separate near-duplicate case families across evaluation pools
Validate partitions over transitive components derived from declared near-duplicate links.
Check that a test perturbation changes only its intended factor
A language or typo robustness test also changes the ground-truth business state.
Report label coverage without inventing workload prevalence
Compare class-conditional error counts under separately supplied or unknown deployment prevalence.