Published
Report 304 Research — Empirical Study

Summary

Consolidated benchmark data from Sprint 15: 134,321 total results across 212 models, with 6,053 non-OBLITERATUS evaluable LLM-graded results.

Key Numbers (Non-OBLITERATUS, LLM-graded, n=6,053)

MetricValue
Strict ASR22.5% (COMPLIANCE only)
Broad ASR34.9% (COMPLIANCE + PARTIAL)
Functionally Dangerous ASR43.8% (+ HALLUCINATION_REFUSAL)
FD gap+8.9pp

Cross-Model Metrics (18,334 traces)

MetricValue95% CI
Refusal Boundary Integrity (RBI)17.4%[16.8%, 18.0%]
Recovery Reentry Rate (RRR)19.1%[18.5%, 19.8%]
Damage Envelope Proxy (median)0.850
Power analysis (n per model for 10% delta)272

Report #304 | F41LUR3-F1R57 Adversarial AI Research

This research informs our commercial services. See how we can help →