weight-perturbation
$
npx mdskill add yogsoth-ai/de-anthropocentric-research-engine/weight-perturbation扰动权重检验 gap 排序稳定性,输出稳定性判定。
SKILL.md
.github/skills/weight-perturbationView on GitHub ↗
---
name: weight-perturbation
description: "SOP: 扰动权重检验 gap 排序稳定性,输出稳定性判定"
version: 1.0.0
category: hypothesis-formation
type: sop
campaign: gap-prioritization
input: "基准权重向量 + gap 评分矩阵(gap × 维度)"
output: "PerturbationReport — 扰动方案、排序变化幅度及稳定性判定"
dependencies:
skills:
- subagent-spawning
---
# Weight Perturbation
扰动权重检验 gap 排序稳定性,输出稳定性判定。
## HARD-GATE
<HARD-GATE>
- 输入权重向量各元素之和必须等于 1.0(允许 ±0.001 误差)
- 评分矩阵行数(gap 数)必须 ≥ 2
- 至少生成 4 个扰动方案(±20% 各维度)
- stability_verdict 必须为 "stable" | "sensitive" | "unstable" 之一
</HARD-GATE>
## Pipeline
1. **前置检查**: 验证权重向量归一化;验证评分矩阵维度与权重向量长度一致
2. **基准排序计算**: 用基准权重对评分矩阵加权求和,得到基准排序
3. **扰动方案生成**: 对每个维度分别施加 +20% 和 -20% 扰动(重新归一化后),生成 2×n 个扰动方案
4. **重新计算排序**: 对每个扰动方案计算新排序
5. **比较变化幅度**: 统计每个方案中排序变化的 gap 数量;计算 Kendall τ 与基准排序的相关性
6. **稳定性判定**: stable(所有方案 τ ≥ 0.8)/ sensitive(任意方案 0.5 ≤ τ < 0.8)/ unstable(任意方案 τ < 0.5)
7. **输出**: 返回 PerturbationReport 对象
## Output Format
```json
{
"baseline_ranking": ["gap_003", "gap_001", "gap_002"],
"perturbation_scenarios": [
{
"scenario_id": "importance_+20%",
"perturbed_weights": { "importance": 0.48, "feasibility": 0.18, "novelty": 0.17, "impact": 0.17 },
"ranking": ["gap_003", "gap_001", "gap_002"],
"kendall_tau": 1.0,
"rank_changes": 0
}
],
"min_kendall_tau": 0.87,
"stability_verdict": "stable",
"sensitive_dimensions": [],
"summary": "稳定性摘要(2-3句)"
}
```
More from yogsoth-ai/de-anthropocentric-research-engine
- abductive-hypothesis-generationStrategy: 面对异常的最佳解释推理
- ablation-brainstormRemove components one by one, observe system changes to reveal hidden dependencies and generate ideas from structural gaps.
- ablation-component-mappingMap system architecture to ablatable units for ablation studies
- ablation-designDesign ablation studies to isolate component contributions in ML systems
- ablation-executionRemove components one by one from a system, record the response/impact of each removal.
- abp-vulnerability-classificationClassify assumptions on 2 axes — load-bearing (how much conclusion depends on it) × vulnerable (how likely to be false). Focuses attention on High-Load × High-Vulnerable quadrant.
- abstraction-extractionExtract abstract principles from concrete domain cases. Strips domain-specific details to reveal transferable mechanisms.
- abstraction-ladderPerform bisociation at multiple abstraction levels
- abstraction-ladderingMove between concrete and abstract framings — 3 levels up (Why?) and 3 levels down (How?) to find the most productive research level.
- abstraction-to-designAbstract biological principle to design principle. Bridge from biology to engineering.