Transparency

Methodology & taxonomy changelog

Every material change to how AIBRAI classifies incidents (sector, severity, harm type, AI involvement, bias basis) or to the intake and publication methodology is logged here. We use semantic versioning: major for breaking taxonomy changes, minor for additive changes, patch for clarifications.

1. v1.3.0

   2026-01-15

   policyplatform

   Granular cookie consent and public data-request flow

   Cookie banner now supports per-category consent (strictly necessary, analytics, marketing) with reject-all as a first-class option. A dedicated /data-request page documents GDPR / Swiss FADP rights and provides a pre-filled contact flow.

   • Added analytics + marketing consent categories, both off by default.
   • Published /data-request with DSAR guidance and identity-verification notes.
   • Published this public methodology & taxonomy changelog.

2. v1.2.0

   2025-11-20

   taxonomy

   Sector enum expanded

   Split the previous 'government' sector into 'government_benefits', 'immigration' and 'law_enforcement' following clustering analysis of Q3 submissions and alignment with the EU AI Act Annex III high-risk categories.

   • Added: government_benefits, immigration, law_enforcement.
   • Deprecated (still accepted, mapped on read): government.
   • Historical entries backfilled where the original narrative unambiguously fit one of the new codes.

3. v1.1.0

   2025-09-02

   taxonomymethodology

   Harm-type taxonomy alignment with NIST AI RMF

   Reworked harm_type values to align with the NIST AI Risk Management Framework 1.0 harm taxonomy so submissions cross-reference cleanly with public research.

   • Renamed: 'financial_loss' → 'economic_harm'.
   • Added: 'psychological_harm', 'loss_of_opportunity', 'representational_harm'.
   • Deprecated: 'other_harm' (replaced by structured combinations).

4. v1.0.0

   2025-06-10

   taxonomymethodologypolicy

   Public v1 of the AIBRAI Bias Registry

   First public release of the registry with the initial taxonomy (sector, severity, harm_type, ai_involvement, bias_basis) and the intake methodology (anonymous submissions accepted, Cloudflare Turnstile for abuse control, salted IP hashing retained for 48h).