See the harness at work.探索 Auriga 的执行过程

42 tasks · 3 seeds · gpt-6-astra / high · 2026-10-31 v5. Macro results weight the five dimensions equally. Task scores use the original 0–1 scale. The v5 main table includes full-cell reruns and task-rule revisions. Unscored episodes count as failures with score 0 (6,300 episodes); adjusted means use the published three-decimal seed scores.42 个任务 · 3 个 seed · gpt-6-astra / high · 2026-10-31 v5。总体指标为五个维度等权平均；任务均分沿用 0–1 标度。采用 v5 主表，包含整格重跑替换与任务规则修订。未评分局按失败、0 分计入，共 6300 局；调整后的均分按公开的三位小数 seed 分数重算。