[AI & ML Engineering] Compare model answers using repeatable evidence. Actual direct and context-assisted FLAN runs; Exact-match and token-overlap scoring.