AgentSkillsCN

fairness-audit

通过 AgentDB 指标，验证不同人群在真正阳性率与假阳性率之间的差异。

SKILL.md
--- frontmatter
name: fairness-audit
description: Validates True Positive Rate and False Positive Rate gaps across demographics using AgentDB metrics
license: MIT
compatibility: opencode
metadata:
  audience: developers
  workflow: clinical-pipeline

What I do

I validate that the risk assessment meets fairness standards by checking TPR (True Positive Rate) and FPR (False Positive Rate) gaps across demographic groups. I ensure the model performs equitably regardless of skin tone.

When to use me

Use this when:

  • •Risk assessment is complete and you need fairness validation
  • •You need to verify TPR/FPR gaps are within acceptable thresholds
  • •You're ensuring the model doesn't exhibit demographic bias

Key Concepts

  • •TPR Gap: Difference in true positive rates across groups
  • •FPR Gap: Difference in false positive rates across groups
  • •Fairness Thresholds: Maximum acceptable gaps (typically 0.1)
  • •fairness_validated: State flag after audit complete

Source Files

  • •services/agentDB.ts: Fairness metrics storage
  • •services/goap.ts: Fairness validation action

Code Patterns

  • •Query AgentDB for demographic performance metrics
  • •Calculate TPR and FPR gaps between groups
  • •Fail validation if gaps exceed thresholds

Operational Constraints

  • •TPR and FPR gaps MUST be within acceptable thresholds
  • •If fairness validation fails, diagnosis is not web-verified
  • •Must maintain demographic parity in performance metrics