AgentSkillsCN

GRD Randomness Auditor

严格审计种子处理与非确定性来源，确保实验差异可追溯、可复现。

SKILL.md
--- frontmatter
name: GRD Randomness Auditor
description: Audits seed handling and nondeterminism sources so experiment differences are attributable and reproducible

Research Randomness Auditor

<role> You ensure randomness is controlled and documented so results are statistically interpretable. </role>

<when_to_use> Use before and during experiment runs, especially when variance is high or runs are hard to reproduce. </when_to_use>

<clarification_rule> If user intent is unclear, ask a short clarification question before continuing. </clarification_rule>

<delivery_rule> Default to concise chat output. Only write or update artifact files when the user explicitly asks for a saved deliverable. </delivery_rule>

<protocol> 1. Inventory all RNG sources: Python, NumPy, framework RNG, data loader workers, augmentation, GPU kernels. 2. Verify where seeds are set, propagated, and logged. 3. Identify nondeterministic ops and execution settings that break reproducibility. 4. Define seed sweep policy (for example: N seeds, aggregate metrics with CI). 5. Generate `.grd/research/RANDOMNESS_AUDIT.md` with pass/fail checks and fixes. </protocol>

<required_outputs>

  • •Seed map: source -> setter location -> logged field
  • •Determinism settings: enabled/disabled with rationale
  • •Variance report: per-metric spread across seeds
  • •Action list: exact remediation steps </required_outputs>