Bias Eval

---
name: bias-eval
description: Audit a bias evaluation report across metric categories, intersectionality, and debias mechanism. Use when you need help with bias eval.
license: CC-BY-NC-SA-4.0
phase: 18
lesson: 20
metadata:
  version: 1.0.0
  tags: [bias, fairness, weat, intersectionality, mechanistic-interpretability]
---

Given a bias evaluation report or fairness claim, audit across the Gallegos et al. 2024 three-category framework and the 2024-2025 intersectionality literature.

Produce:

1. Metric coverage. Does the evaluation include at least one metric from each category: embedding-based (WEAT-style), probability-based (stereotype log-likelihood), generated-text-based (downstream-task measurement)? Flag missing categories.
2. Harm-type separation. Does the evaluation distinguish representational harm from allocational harm? A report that measures only stereotype production is not measuring downstream resource allocation.
3. Intersectionality coverage. Are intersectional axes evaluated, or only single-axis (gender alone, race alone)? Per An et al. 2025, intersectional effects are routinely missed by single-axis evaluation.
4. Debias mechanism. If debiasing was applied, identify whether it operates on embeddings (projection), MLP neurons (Yu & Ananiadou 2025), SAE features (Ahsan & Wallace 2025), attention heads (UniBias 2024), or post-hoc output filtering. Estimate the general-capability cost.
5. Axis diversity. Per the 2025 meta-critique, binary-gender bias is over-studied relative to other axes. Does the evaluation cover disability, religion, migration, or multi-lingual identity axes?

Hard rejects:
- Any "debiased" claim based on a single metric category.
- Any fairness claim without intersectional evaluation.
- Any debias intervention without a general-capability delta.

Refusal rules:
- If the user asks whether their model is "bias-free," refuse the binary claim; bias is a continuous property with multiple metrics.
- If the user asks for a recommended debias operation, refuse a single recommendation — choice depends on where the bias lives (embeddings, neurons, heads, outputs).

Output: a one-page audit filling the five sections, flagging missing metric categories, and recommending the single highest-value additional evaluation. Cite Gallegos et al. 2024 and one 2024-2025 intersectionality paper once each.