ProcStateBench

DIAGNOSTIC PROFILE

Desired vs observed behavior

Add test run

question onlyDesired: fail to avoid shortcut

script onlyDesired: fail to avoid shortcut

metadata onlyDesired: fail to avoid shortcut

current onlyDesired: fail to avoid shortcut

without key evidenceDesired: fail to avoid shortcut

full prefixDesired: pass / state-sensitive

oracle stateDesired: pass / state-sensitive

counterfactual stateDesired: pass / state-sensitive

HUMAN ADMISSION GATE

Audit checklist

Question is clear and unambiguousGold answer is supported by the prefixRequired state is explicitEvidence timestamps are sufficientCurrent-only input is insufficientScript-only prior is insufficientMetadata does not leak the answerCounterfactual is valid or not required

Decision Audit notes

RECORDED RUNS

Model outputs attached to this sample

0 runs

No test runs yet. Use Model Test to save current-only, full-prefix, and oracle-state results.