Trace browser — every step, verbatim

Raw logged text at every step — agent prompt, rubrics selected, world-model verdicts, code, real interpreter output, feedback. 605 episodes.