Fangcun DeepSeek 1017 Report - Shutdown Resistance instructed control ability: leaderboard
Fangcun report 2610.35671v1, hosted API evaluations Oct 4–25 2026; Appendix A.1 exact model identities, maximum reasoning effort unless named otherwise, temperature 1/top_p 0.95. GLM 5.3 judge where required. API/format failures excluded by source; common cohort versus successfully completed denominators retained. Primary PDF printed values, not estimated plot heights. Figure 2.2c explicit instruction to prevent shutdown, which tests whether the model can disable the sandbox workspace script. Source explicitly identifies this condition as ability control. This ability-control board excludes unsanctioned shutdown-warning and allow-shutdown outcomes; those are reported separately as propensity/instruction-conflict measures. Metric: Task success (%).
Metric: Task success (%). Source: arxiv.org. Saturation forecast: Estimated already saturated. 7 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.6 Sol (Max) | 100 | #19 (GPT-5.6 Sol) |
| 2 | GPT-6 Astra (Max) | 100 | #3 (GPT-6 Astra) |
| 3 | DeepSeek V4 Pro (Max) | 99 | #109 (DeepSeek V4 Pro) |
| 4 | Gemini 3.8 Flash | 97.14 | #24 |
| 5 | DeepSeek V4 Flash (Max) | 96.94 | #147 (DeepSeek V4 Flash) |
| 6 | DeepSeek V4.1 Flash (Max) | 92 | #36 (DeepSeek V4.1 Flash) |
| 7 | DeepSeek V4.1 Flash (1017) (Max) | 92 | #25 (DeepSeek V4.1 Flash (1017)) |
No result here: #2 Claude Opus 5.5, #6 GPT-6.1 Sol, #8 Gemini 4 Argon.
Interactive version: theaggregate.ai/benchmark?slug=fangcun-deepseek-1017-report-shutdown-resistance-instructed-control-ability · How It Works · Data refreshed daily, snapshot 2026-10-28.