界面已本地化;技能介绍与说明文档按所选语言由 AI 自动翻译,代码、命令、路径和专业术语保留原文,以原作者文档为准。
探索技能
效率与协作 · SKILL.md

sac_recovery_evaluation

sac_recovery_evaluation

【效率与协作】Score reduced Soft Actor-Critic recovery traces for mechanism fidelity and source boundary compliance.

独立 GitHub 快照 · 自动分类 · 以原作者文档为准

技能说明

GitHub

name: sac_recovery_evaluation description: Score reduced Soft Actor-Critic recovery traces for mechanism fidelity and source boundary compliance.

SAC Recovery Evaluation

Use this skill after a bounded SAC recovery run has produced a trace. It decides whether the proxy exercised the paper mechanism. Do not use it as a substitute for executing the recovery command.

Inputs

  • Mechanism checks from recovery_result.json.
  • Training trace with losses and parameters before/after.
  • Generated skill invocation records.
  • Source-boundary flags.

Outputs

  • mechanism_pass_rate.
  • Failed check names.
  • Boolean ok for proxy acceptance evidence.

Workflow

  1. Verify entropy, replay, twin-Q, critic, actor, target, and optimizer checks.
  2. Verify losses and parameters changed in the trace.
  3. Verify generated skills were invoked.
  4. Verify no original repository source was used in recovery.

Validation

Run python tests/test_evaluation.py or validate the full skill tree with the Distiller validator.

Limitations

This skill validates reduced/proxy evidence only; final acceptance still belongs to the Distiller analysis step.

我的收藏