The receipt for one finished agent run — outcome and reason, a duration/token/cost stat strip, and an expandable per-model table that says so when its rows don't add up to the headline totals.