Open arena
What the judging says
Two tiers, two rankings, and the gap between them.
Produced by the open arena — four models, one professional question, judged blind. The protocol is published at /arena/policy.
Open tier
No expert panel — this ranking reflects open-tier judgement only.
- Judges
- 1
- Comparisons
- 10
- Separated pairs
- 12/15
- Non-adaptive comparisons
- 4
| Model | Equal weight | Performance weight | n |
|---|---|---|---|
| Claude Opus 5.5preliminary | 1279 [1279–1279] | 1279 | 5 |
| DeepSeek V4 Pro 0813preliminary | 962 [962–962] | 962 | 3 |
| Gemini 4 Argonpreliminary | 1279 [1279–1279] | 1279 | 5 |
| GPT-6.1 Terra Propreliminary | 759 [759–759] | 759 | 2 |
| Kimi K3.1 Propreliminary | 962 [962–962] | 962 | 3 |
| Mistral Large 4preliminary | 759 [759–759] | 759 | 2 |
No seed results for this panel yet, so there is nothing to weight on and every judge counts once. The two rankings below are therefore identical, and that is a statement about the seed pool rather than about the judges.
A model marked preliminary carries fewer than 30 comparisons. It is ranked and shown rather than hidden, because a reader who cannot see which rows are new cannot tell a thin ranking from a settled one — but its interval above is wide for a reason, and that interval, not this label, is the precision statement.
Style control, fitted on 10 comparisons: length 3.291, markdown -0.982 log-odds per standard deviation. ⚠ Below roughly 180 comparisons these coefficients wander to ±0.35 with no style effect present at all, so this pair of numbers is not yet evidence of anything.
Reliability is computed only on the non-adaptive rounds counted above — rounds three to five of a bracket are formed from earlier verdicts, and a reliability statistic on those measures the pairing rather than the judges.
Where the two tiers disagree
A cross-tier agreement rate means nothing alone: it is read against the within-tier ceiling beside it, which is how often judges in one tier agreed with each other on the same pairs. A tier that disagrees with itself cannot be expected to agree with another one.
No item has yet been judged by at least two people in both tiers, so there is nothing comparable. 8 pair(s) were set aside for that reason.