SafePyramid - L0 Rule Understanding: leaderboard

Metric: Rule matching rate (%) on the 1,000 L0 cases (independent decisive and distractor 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 July 2028. 12 models tracked.

Top models

#ModelScore
1Claude Opus 4.7 (Max)73.8
2DeepSeek V4 Pro (Max)70.8
3GPT-5.5 (xHigh)70.7
4Grok 4.3 (High)70.6
5Qwen 3.6 Max (Preview) (Thinking)68.9
6Seed 2.0 Pro (High)63.1
7Gemini 3.5 Flash (High)61.4
8Hy3-preview (High)60.6
9gpt-oss-safeguard-20B55.7

Interactive version: theaggregate.ai/benchmark?slug=safepyramid-l0-rule-understanding · How It Works · Data refreshed daily, snapshot 2026-09-29.