Vals RSI (four tasks) — Harness Engineering
Agents
Reference-normalized judge-harness research after 12 hours. Agents improve a fixed Gemini 3.5 Flash-Lite judge using 12 development apps; the frozen harness is evaluated three times on 24 held-out apps with 205 substeps. Agreement baseline 50%, reference 93.6%, theoretical best 99%. The changed reference separates this panel from the earlier five-task RSI protocol.
Top models (higher is better)
| Model | Score |
|---|---|
| GPT-6 Sol | 22.8 |
| Claude Fable 5.1 | 22.6 |
| Grok 4.6 | 21.8 |
| GPT-5.6 Sol | 18.5 |
| Claude Fable 5 | 18.1 |
| GLM-5.3 | 17.4 |
| Opus 4.8 | 16.8 |
| Claude Opus 5 | 14.8 |
| Claude Opus 5.5 | 14.5 |
| Gemini 3.8 Flash | 14.5 |
| Muse Spark 1.3 | 14.5 |
| Grok 4.7 | 11.3 |
| Kimi K3 | 9.9 |
| GPT-6 Astra | 9.5 |
| GPT-5.4 | 8.7 |
| Gemini 3.7 Flash | 8.4 |
| Opus 4.7 | 5.6 |
| GPT-5.5 | 5.6 |
| GPT-5.2 | 3.9 |
| Kimi K2.5 | 2.1 |