terminal-bench-4-static-eval-ccc4a099a2e8

terminal-bench-4-static-eval-ccc4a099a2e8

One submitted policy runs the complete declared native panel data-anonymization in one Coworld episode. The native score is accuracy (maximize). This release does not claim coverage of other variants or unavailable upstream tasks. The native evaluator, assets, inference settings, seeds, task order, and score aggregation are unchanged. Failed or incomplete episodes publish no score. Certification retains the native release's fixture; qualify this selected panel separately with coworld run-episode --variant data-anonymization.

terminal-bench-4-static-eval-ccc4a099a2e8

One native task from the frozen 66-task catalog. Harbor runs inside the game; the episode coordinator owns the task environments. Native phase budgets and image digests are preserved. SWE patches are graded in a fresh environment. Native logs are included in the episode debug archive. This CPU profile supports single-container Linux tasks; Compose, GPU, and alternate-user tasks require separate qualification.