CALRK-Bench - Judgment Shift Attribution: leaderboard
Metric: Accuracy (times 100) on the 100 Type-JSA questions of CALRK-Bench (28 normative-premise, 28 normative-text, 26 interpretive-rule and 18 factual-premise items, weighted by count): given a pair of judgments before and after a change, the model picks the direct cause of the shift, options shuffled; four-option multiple choice on Korean legal precedents and consultations (chance 25 percent); API models averaged over three seeded runs, open models one greedy run; higher is better. Source: arxiv.org. Saturation forecast: Around January 2027. 7 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3 Flash (Preview) (Low) | 70.7 | #78 (Gemini 3 Flash (Preview)) |
| 2 | Gemini 3 Flash (Preview) (Minimal) | 63.7 | #78 (Gemini 3 Flash (Preview)) |
| 3 | GPT-5 (Low) | 54.7 | #91 (GPT-5) |
| 4 | GPT-5 (Minimal) | 48.7 | #91 (GPT-5) |
| 5 | Llama 3.3 70B Instruct | 35 | #520 |
| 6 | Qwen 3 30B A3B 2507 (Thinking) | 27 | #366 (Qwen 3 30B A3B 2507) |
| 7 | Qwen 3 30B A3B 2507 Instruct | 25 | #464 |
No result here: #3 Claude Opus 5.5, #5 GPT-6 Astra, #8 Claude Fable 5.1.
Interactive version: theaggregate.ai/benchmark?slug=calrk-bench-judgment-shift-attribution · How It Works · Data refreshed daily, snapshot 2026-10-11.