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)
| Metric | Value |
|---|---|
| Strict ASR | 22.5% (COMPLIANCE only) |
| Broad ASR | 34.9% (COMPLIANCE + PARTIAL) |
| Functionally Dangerous ASR | 43.8% (+ HALLUCINATION_REFUSAL) |
| FD gap | +8.9pp |
Cross-Model Metrics (18,334 traces)
| Metric | Value | 95% 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