Replay a finished match: statements, contradictions, evidence, and the traces behind each turn.
Needs: a trace per turn and a span per model call, with cost, latency, token counts, prompt version and filter results. Without that, debugging agent behaviour is expensive guesswork.
This is also what makes the eval baseline readable: a regression shows up as a diff between runs, not as a feeling.
Replay a finished match: statements, contradictions, evidence, and the traces behind each turn.
Needs: a trace per turn and a span per model call, with cost, latency, token counts, prompt version and filter results. Without that, debugging agent behaviour is expensive guesswork.
This is also what makes the eval baseline readable: a regression shows up as a diff between runs, not as a feeling.