- OpenAPI contract implemented end-to-end (
openapi/aibf.yaml) - Feature model splitting merit vs protected-proxy signals
- Linear reference model + zero-baseline (SHAP) attributions
- Bias score, thresholds, plain-language explanations, flagging
- HR feedback + human-in-the-loop retraining and threshold recalibration
- Synthetic, deterministic training/demo data (no real PII)
- Test suite + runnable pipeline demo
- Pluggable non-linear reference models with kernel/tree SHAP.
- Counterfactual explanations ("would not have been flagged if …").
- Proxy-variable detection for name/zip-code style leakage.
- Map bias scores to the four-fifths (80%) rule and impact-ratio reporting.
- Exportable, machine-readable audit reports (JSON + human summary).
- Alignment notes for NYC Local Law 144 and EU AI Act Article 10 obligations.
- Adapters to consume real ATS outputs (Greenhouse/Lever/iCIMS-style APIs).
- Consented, self-reported demographic ingestion (replacing inference).
- Reviewer UI for the flagged-decision feed.
- Scheduled re-audits with model-drift and protected-class coverage checks.
- Dashboards for population-level disparate-impact trends over time.
- Making legal determinations of discrimination.
- Shipping or storing real candidate PII in the open-source core.