Loading runs…
Explore 1,482 traces contributing to Figure 1 across nine agents. The original repeated-run cohorts contribute 1,050 traces; the attempt-rate bars use another 90 Claude Fable 5.1 traces across 37 tasks and 342 GPT-6 Sol traces across 50 tasks. Those two attempt-rate evidence sets include mixed campaigns and effort settings, and differ from the success-rate cohorts. Only selected Figure 1 evidence is published here.
Loading runs…