Fresh attempts recorded for this page on 11 Oct 2026 on Fireworks, both arms loaded from saved state into a fresh session (the same route as the evaluation), with the same agent, prompt, tool and limits, at temperature 1.0. Returned rows are shown up to 20; scores were computed on the full result. Every message is shown unedited: the model's reasoning, each query it ran, and exactly what the tool returned. Long reasoning is folded, not cut — "show all" expands it.