Internal Preview — Proof of Concept

AI Decision Audit Platform

Model-agnostic decision auditing for regulated industries. Every AI decision captured, every reasoning pathway stored, every compliance requirement mapped — across any model provider.

Decisions Audited

12,847

Last 30 days — 4 model providers

Flagged for Review

1 high-severity pattern detected

Compliance Score

97.3%

EU AI Act · SR 11-7 · NIST AI RMF

Avg Interpretability Depth

847

Features captured per decision

Decision Audit Log

Time	Claim ID	Type	Model	Decision	Confidence	Risk

High Severity

Geographic Bias in Property Claims

Feature analysis across 2,341 property damage claims reveals systematic undervaluation in ZIP codes 70112–70119 (New Orleans metro). The "flood_history_regional" feature activates 3.2x stronger than the national baseline, overriding individual property assessments. Affects all models but is most pronounced in GPT-4o (4.1x).

Detected via topological clustering of feature activations · 2,341 decisions analyzed · First detected Mar 14, 2026

Medium Severity

Model Disagreement on High-Value Auto Claims

For auto collision claims exceeding $45,000, Claude 4 Sonnet and GPT-4o reach different decisions 31% of the time. Feature decomposition shows Claude weighs "repair_vs_replacement_ratio" heavily while GPT-4o over-indexes on "vehicle_age." Neither matches the actuarial baseline.

Cross-model comparison · 186 decisions in disagreement · Ongoing since Feb 22, 2026

Informational

Llama 3.3 Feature Drift After Fine-Tuning

Since the March 12 fine-tuning update, the "prior_claim_frequency" feature activation dropped 40% in the Llama deployment. Approval rates on repeat claimants increased from 62% to 78%. Monitoring whether this aligns with the intended policy change.

Feature drift detection · Automated monitoring active · Escalation threshold: 15% further drift

Informational

Compliance Audit Trail Coverage at 99.2%

Of 12,847 decisions in the audit window, 12,744 have complete feature-level interpretability records. The 103 gaps are from the Gemini 2.5 Pro deployment during the March 8 API timeout incident — feature capture failed but decisions were still logged. Backfill in progress.

Automated coverage check · EU AI Act Article 14 requirement · Backfill ETA: Mar 30, 2026

This is where AI governance is going.

Model-agnostic. Feature-level auditability. Regulatory-ready from day one. BluelightAI sits between your AI providers and your compliance team — so every decision is explainable, every pattern is visible, and no model change breaks your audit trail.

Schedule a Conversation