Evidence provenance
Recording and
evaluation provenance.
The featured BlendIngredients recording comes from historical job 20261009T221300Z-0daa0ac68c, Opus 5.5, attempt two of a two-attempt pilot. Its trusted upstream predicate reported success sustained for 21 control steps after 1,526 steps. It is an observed pilot result, not a five-trial RoboSWE-ADAPT leaderboard entry.
The video runs on the recorded simulation clock at 20 control steps per second. The overlay is aligned to each model decision’s preceding observation step. Thinking pauses are removed; multiple decisions without robot motion may occupy the same video timestamp.
Action rationales and submitted code are exact fields from the recorded decision events. They are not hidden reasoning, editorial reconstructions, or a claim that the rationale accurately explains the physical result. The overview is a spectator view; the robot camera is an observation view.
Download the curated evidence and SHA-256 source hashes. The export omits provider requests, credentials, private runtime metadata, and raw hidden reasoning.
Task selection
The development catalogue contains 83 selections, representing 85 native tasks when the two proposed combinations are expanded. The prospective public pilot contains 11 runnable task candidates; each scene requires admission before evaluation. Selection emphasized longer tasks with six or more subtasks and diverse skills, supplemented with shorter calibration tasks. Runtime admission and exclusions are recorded separately before release scoring.