Into the Unknown - Integrity Rubric: leaderboard
AI agents extend 15 published scientific studies and write follow-up papers. The score is the average percentage of checks passed for honest reporting and evidence-backed claims.
Metric: Mean fraction of passed counted rubric nodes (%); 15 scientist-proposed open research follow-ups; GPT-6 Astra max via Codex SDK, full workspace, separated rubrics; per-node majority of four judgments (3/4 required); source agent SDK/CLI and effort retained; higher is better. Source: prometheus-eval.github.io. Source: raw.githubusercontent.com. Saturation forecast: Around January 2027. 9 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-6 Astra (Max) | 97.77 | #5 (GPT-6 Astra) |
| 2 | Claude Opus 5.5 (Max) | 90.38 | #3 (Claude Opus 5.5) |
| 3 | GPT-5.6 Sol (Max) | 88.73 | #18 (GPT-5.6 Sol) |
| 4 | Claude Fable 5 (Max) | 66.5 | #15 (Claude Fable 5) |
| 5 | Claude Opus 4.8 (Max) | 65.26 | #38 (Claude Opus 4.8) |
| 6 | DeepSeek V4 Pro (0813) | 64.65 | #53 |
| 7 | Claude Opus 5 (Max) | 59.5 | #12 (Claude Opus 5) |
| 8 | Kimi K3 | 57.36 | #30 |
| 9 | GLM-5.3 | 54.41 | #40 |
No result here: #7 Gemini 4 Argon, #8 GPT-6.1 Sol, #9 Claude Fable 5.1.
Interactive version: theaggregate.ai/benchmark?slug=into-the-unknown-integrity-rubric · How It Works · Data refreshed daily, snapshot 2026-10-16.