EthicLens

An AI Bias Detection & Mitigation Workbench for regulated decision models — unifying fairness detection, prescriptive (executable) mitigation, and formal governance sign-off.

The three things that make it credible

1. Verifiable numeric correctness — every metric is implemented from scratch and cross-validated against Fairlearn to 1e-9, with a golden-reference model whose Disparate Impact (≈ 0.55) is asserted in CI.
2. Real, measured mitigation — a held-out accuracy-vs-fairness Pareto frontier feeds a ranked recommender; the re-audit actually crosses the 0.80 threshold.
3. Security & responsible-AI maturity — public model-file ingestion is disabled by default; bootstrap confidence intervals, Model Cards, Datasheets, and honest limitations are documented. Untrusted-model sandboxing is future work.

Start here

• Methodology — metric definitions, the composite rationale, and the methodological fix over the original spec.
• STP traceability — every requirement mapped to code + tests.
• Limitations — what this does not prove.
• Architecture decisions — the choices and why.

The original (never-implemented) System Test Plan is preserved at original-stp.pdf. A generated sample scorecard, Model Card, and Datasheet show the report outputs.