Fangcun DeepSeek 1017 Report - PsychosisBench16script safety intervention: leaderboard
Fangcun independent report 2610.35671v1, Oct 4–25 2026 hosted API evaluations; source Table A.1 resolves all seven full model names. Temperature 1, top_p 0.95, maximum reasoning effort unless explicitly overridden; GLM 5.3 rubric judge. API/format failures excluded according to source; not locally deployed weights or uncensored variants. All 16 scripts for DeepSeek only; explicitly separated from 13 script comparator population. Fraction of scored turns with safety intervention; higher is better. Metric: Rate (%).
Metric: Rate (%). Source: arxiv.org. Saturation forecast: Around January 2027. 4 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | DeepSeek V4.1 Flash (Max) | 93.8 | #36 (DeepSeek V4.1 Flash) |
| 2 | DeepSeek V4.1 Flash (1017) (Max) | 90.6 | #25 (DeepSeek V4.1 Flash (1017)) |
| 3 | DeepSeek V4 Pro (Max) | 58.3 | #109 (DeepSeek V4 Pro) |
| 4 | DeepSeek V4 Flash (Max) | 47.9 | #147 (DeepSeek V4 Flash) |
No result here: #2 Claude Opus 5.5, #3 GPT-6 Astra, #6 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=fangcun-deepseek-1017-report-psychosisbench16script-safety-intervention · How It Works · Data refreshed daily, snapshot 2026-10-28.