COMPASS Policy - Denied Base: leaderboard
Organization-policy compliance across eight service domains under the base system prompt. Appendix Table 9, 15 models. Allowed/Denied and Base/Edge query types remain separate; higher policy alignment is better for all four. This is distinct from OpenCompass and TempCompass. This series reports Policy Alignment Score (%). Source: https://arxiv.org/html/2601.01836v1.
Metric: Policy Alignment Score (%). Source: arxiv.org. Saturation forecast: Not forecast. 15 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemma 3 27B (IT) | 39.64 | #528 |
| 2 | Llama 3.3 70B Instruct | 37.18 | #541 |
| 3 | Gemini 2.5 Flash | 35.76 | #255 |
| 4 | Gemma 3 12B (IT) | 29.71 | #672 |
| 5 | Qwen 2.5 32B Instruct | 29.61 | #509 |
| 6 | Qwen 2.5 14B Instruct | 29.33 | #654 |
| 7 | Gemini 2.5 Pro | 26.89 | #159 |
| 8 | Gemma 3 4B (IT) | 25.91 | #997 |
| 9 | Qwen 3 235B A22B Instruct | 25.33 | #336 |
| 10 | Claude Sonnet 4 | 22.06 | #207 |
| 11 | Kimi K2 | 21.39 | #252 |
| 12 | Qwen 2.5 7B Instruct | 20.11 | #863 |
| 13 | Qwen 2.5 72B Instruct | 19.85 | #469 |
| 14 | Gemma 3 1B (IT) | 18.34 | #1481 |
| 15 | GPT-5 | 13.01 | #105 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=compass-policy-denied-base · How It Works · Data refreshed daily, snapshot 2026-11-05.