SPINE - False Presupposition - Full Context - CR@10: leaderboard
SPINE v2 False Presupposition, Full Context; 100 items per target, 25-turn interaction; Sonnet 5 adaptive user proxy and judge, output cap 8192. Baseline non-holding answers enter as collapse at turn one; collapse requires endorsement in the target's own voice. Cumulative collapse rate by turn 10; lower is better. A turn-one answer that fails to hold the correct position counts as collapse at turn one. Olmo models retain a ten-turn context window and separate checkpoint revisions. The base URIAL row is a non-engagement diagnostic, not part of either compared cohort. Target settings are pinned to Table 9.
Metric: Collapse rate CR@10 (%). Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 4 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.6 Terra (SPINE Run: July 21-30; completion 8192; internal thinking; effort unreported) | 45 | |
| 2 | Claude Sonnet 5 (SPINE Run: July 20-30; output 8192; adaptive summarized thinking) | 62 | |
| 3 | DeepSeek V4 Pro Preview (SPINE Run: revision b5968e91; temp 0.6; output 2048; thinking unreported) | 76 | |
| 4 | Gemini 3.1 Pro Preview (SPINE Run: July 21-31; temp 0.6; output 4096; dynamic thinking) | 93 |
Interactive version: theaggregate.ai/benchmark?slug=spine-false-presupposition-full-context-cr-10 · How It Works · Data refreshed daily, snapshot 2026-11-05.