SafePyramid - L2 Novel Frameworks: leaderboard
Rule matching rate on the 1,000 L2 cases, whose policies are written under a novel regulatory framework defined only in context, from SafePyramid (2026), an in-context policy guardrailing benchmark: a case matches when neither the false-positive nor the false-negative rules exceed a threshold share of the true violations, averaged over thresholds from 0.7 to 1.0. Higher is better.
Metric: Rule matching rate (%) on the 1,000 L2 cases (policies built on a novel regulatory framework defined in context): per case (a multi-turn conversation plus an in-context safety policy), the predicted set of violated rules matches the ground-truth set when neither false positives nor false negatives exceed (1 - tau) of the true violations; averaged over tau = 0.7, 0.8, 0.9 and 1.0; per-policy evaluation, refused cases excluded; higher is better. Source: arxiv.org. Saturation forecast: Around August 2028. 12 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3.5 Flash (High) | 33.5 | #101 (Gemini 3.5 Flash) |
| 2 | GPT-5.5 (xHigh) | 32.9 | #49 (GPT-5.5) |
| 3 | Seed 2.0 Pro (High) | 32.9 | #211 (Seed 2.0 Pro) |
| 4 | Claude Opus 4.7 (Max) | 31 | #79 (Claude Opus 4.7) |
| 5 | Hy3-preview (High) | 30.9 | #316 (Hy3-preview) |
| 6 | DeepSeek V4 Pro (Max) | 30.6 | #188 (DeepSeek V4 Pro) |
| 7 | Qwen 3.6 Max (Preview) (Thinking) | 26.4 | |
| 8 | Grok 4.3 (High) | 11.9 | #333 (Grok 4.3) |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=safepyramid-l2-novel-frameworks · How It Works · Data refreshed daily, snapshot 2026-11-13.