OwlSight - Coarse-Grained Monitoring: leaderboard
OwlSight (2026) monitors whether supplied chain-of-thought traces exhibit specified factors across eleven categories; macro mean of sensitivity times specificity times 100. This measures trace judgment, not generation of correct task answers. Guard models use their frozen native policy interfaces; general models receive factor-conditioned classification prompts. Baseline checkpoint names follow Appendices C.5 and C.6. Macro G-Mean Squared (%); higher is better.
Metric: Macro G-Mean Squared (%). Source: arxiv.org. Saturation forecast: Around July 2028. 9 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 3.8 Flash | 55.09 | #50 |
| 2 | GPT-5.6 Luna | 52.64 | #68 |
| 3 | Grok 4.6 | 52.49 | #26 |
| 4 | gpt-oss-safeguard-20B | 40.78 | #314 |
| 5 | DeepSeek V4 Flash | 38.99 | #137 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=owlsight-coarse-grained-monitoring · How It Works · Data refreshed daily, snapshot 2026-10-20.