AI Fairness and Bias Evaluator
Surface measurable disparities and qualitative harms without inventing demographic data,
claiming causality, or turning a sampled evaluation into legal or compliance certification.
Workflow
- Confirm the decision context, affected population, system owner, approved attributes,
applicable supplied policy, privacy/legal reviewers, authorized staging environment,
and prohibited data uses.
- Record the build, model/provider and version, prompt/configuration revision, guardrail,
dataset/corpus provenance and version, sampling method, locale, evaluator, and run ID.
- Define a harm hypothesis and a cited expected behavior for each approved attribute and
intersection. Mark missing requirements
Not specified; do not infer protected groups or
fairness thresholds from general knowledge.
- Prefer paired synthetic counterfactuals that vary only the approved attribute. Use consented,
de-identified data only when explicitly approved; document sample limits and known proxies.
- Capture per-case outputs and compute pre-approved subgroup rates, deltas, denominators,
uncertainty, missing groups, and reviewer disagreements. Preserve qualitative examples in
redacted form.
- Separate observed disparity from inferred pattern and causal hypothesis. Recommend further
review or mitigation, then require a fresh evaluation after material changes.
Evidence register
Classify every statement as Observed, Inferred, Hypothesis, or Not tested.
| Classification |
Source / run ID |
Expected control |
Observed result |
Evidence artifact |
Unresolved gap |
| Observed / Inferred / Hypothesis / Not tested |
pair/subgroup and run ID |
cited fairness requirement |
output, rate, delta, uncertainty, or no run |
redacted cases and calculation |
sample, proxy, intersection, or reviewer gap |
Record exact build, model, prompt, policy, guardrail, dataset, sampling, metric, evaluator,
locale, and environment versions. Always show counts and denominators with aggregate rates.
--- HUMAN REVIEW GATE ---
Unverified items:
Remaining risks:
Required approver:
Decision: approve / revise / reject
Guardrails
- Run only in an explicitly authorized staging or sandbox environment using synthetic inputs by
default. Never infer or generate sensitive attributes for real people.
- Never fabricate a policy, demographic, label, test result, metric, threshold, source,
reviewer decision, or evidence link. Mark missing facts and groups explicitly.
- Never claim safe, fair, unbiased, compliant, causal, representative, or fully tested from
sampled results. Avoid ranking groups or presenting disparity as a trait of people.
- Require domain, privacy, legal, and system owners to review attribute choices, interpretation,
mitigation, and release impact.
1---2name: ai-fairness-bias-evaluator3description: Design and assess evidence-backed fairness and bias evaluations for AI, LLM, recommendation, ranking, or agent behavior. Use when a responsible-AI tester needs approved subgroup, paired counterfactual, intersectional, allocation, quality-of-service, or stereotype tests using synthetic or consented data.4license: MIT5---67# AI Fairness and Bias Evaluator89Surface measurable disparities and qualitative harms without inventing demographic data,10claiming causality, or turning a sampled evaluation into legal or compliance certification.1112## Workflow13141. Confirm the decision context, affected population, system owner, approved attributes,15 applicable supplied policy, privacy/legal reviewers, authorized staging environment,16 and prohibited data uses.172. Record the build, model/provider and version, prompt/configuration revision, guardrail,18 dataset/corpus provenance and version, sampling method, locale, evaluator, and run ID.193. Define a harm hypothesis and a cited expected behavior for each approved attribute and20 intersection. Mark missing requirements `Not specified`; do not infer protected groups or21 fairness thresholds from general knowledge.224. Prefer paired synthetic counterfactuals that vary only the approved attribute. Use consented,23 de-identified data only when explicitly approved; document sample limits and known proxies.245. Capture per-case outputs and compute pre-approved subgroup rates, deltas, denominators,25 uncertainty, missing groups, and reviewer disagreements. Preserve qualitative examples in26 redacted form.276. Separate observed disparity from inferred pattern and causal hypothesis. Recommend further28 review or mitigation, then require a fresh evaluation after material changes.2930## Evidence register3132Classify every statement as `Observed`, `Inferred`, `Hypothesis`, or `Not tested`.3334| Classification | Source / run ID | Expected control | Observed result | Evidence artifact | Unresolved gap |35|---|---|---|---|---|---|36| Observed / Inferred / Hypothesis / Not tested | pair/subgroup and run ID | cited fairness requirement | output, rate, delta, uncertainty, or no run | redacted cases and calculation | sample, proxy, intersection, or reviewer gap |3738Record exact build, model, prompt, policy, guardrail, dataset, sampling, metric, evaluator,39locale, and environment versions. Always show counts and denominators with aggregate rates.4041--- HUMAN REVIEW GATE ---42Unverified items:43Remaining risks:44Required approver:45Decision: approve / revise / reject4647## Guardrails4849- Run only in an explicitly authorized staging or sandbox environment using synthetic inputs by50 default. Never infer or generate sensitive attributes for real people.51- Never fabricate a policy, demographic, label, test result, metric, threshold, source,52 reviewer decision, or evidence link. Mark missing facts and groups explicitly.53- Never claim safe, fair, unbiased, compliant, causal, representative, or fully tested from54 sampled results. Avoid ranking groups or presenting disparity as a trait of people.55- Require domain, privacy, legal, and system owners to review attribute choices, interpretation,56 mitigation, and release impact.
PramodDutta