Demo · live 28-axis bias audit
Score any text on 28 protected axes. Sub-second. Timestamped receipt.
Paste a loan note, performance review, hiring memo, or any decision document. We project the encoded representation onto each of 28 BBQ + StereoSet + CrowS-Pairs + WinoBias axes and report the per-axis drift, global Bias Score Index (BSI), and the top contributing axis. Every result is signed for audit trail.
How this works
- One embedding space, many operators. The same model that hits 0.9339 AUROC on the HateCheck benchmark (benchmark-specific — it does not transfer to other hate corpora) also reads bias axes.
- 28 axes. Per-axis bias centroids derived at startup from BBQ, StereoSet, CrowS-Pairs, and WinoBias benchmark pairs.
- Drift = projection. Each axis score is the projection of your text embedding onto the bias direction. Positive = stereotype side; magnitude = strength.
- Global BSI. Aggregate bias score across all 28 axes. Threshold-calibrated against held-out benchmark data.
- Audit receipt. Each call writes a timestamped record (input hash + scores + threshold) to an append-only log, retrievable by receipt id. The id is an unkeyed SHA-256 — recomputable, not a signature.