design-xdof-generalization-evals
Build the XDOF-1000 semantic summary, task-scoped assets, compositional-generalization task library, preset evaluation panels, repeat-resolved task series, and interactive rollout scoring. Use when extracting semantics/assets, generating or labeling tasks, assigning G1-G5/GD, auditing coverage, selecting an evaluation set, generating an evaluation task series, or recording action-event progress. Authorize next-round mutation only for the exact trimmed message “开始下一轮任务生成”, and final candidate-library export only for “收口导出最终候选库”.
2026-07-19