Diagnoses incidents from operational evidence and proposes recovery; does not change live systems.
Evaluates observed product outcomes against a prior hypothesis; does not run experiments or change user treatment.
