SafePyramid - L1 Rule Dependencies: leaderboard

Metric: Rule matching rate (%) on the 1,000 L1 cases (policies that add exception and conditional rules): 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 April 2027. 12 models tracked.

Top models

#ModelScore
1GPT-5.5 (xHigh)56.8
2Hy3-preview (High)55.9
3DeepSeek V4 Pro (Max)54.5
4Seed 2.0 Pro (High)54.1
5Gemini 3.5 Flash (High)53.9
6Claude Opus 4.7 (Max)51.1
7Qwen 3.6 Max (Preview) (Thinking)46.5
8Grok 4.3 (High)39.4
9gpt-oss-safeguard-20B8.8

Interactive version: theaggregate.ai/benchmark?slug=safepyramid-l1-rule-dependencies · How It Works · Data refreshed daily, snapshot 2026-09-29.