Interpretive Canons (BVerfG): leaderboard
A sentence-level benchmark (Ringe, ICML 2026 AI4Law Workshop) that classifies the legal reasoning in decisions of the German Federal Constitutional Court by the interpretive canons of Larenz in the tradition of Savigny, with labels from a legal expert. The board is the mean positive-class F1 in percent over seven binary subtasks (two reading criteria, the argument gate and four canons) on held-out test splits, with one expert-written prompt for every model; higher is better.
Metric: Mean positive-class F1 over the seven binary subtasks (%; sentence-level classification of reading criteria, argument support and the four interpretive canons of Larenz in German Federal Constitutional Court decisions; expert-written prompt with worked examples (the same for every model), reasoning models at temperature 1.0; decision-disjoint held-out test split (60% of each subtask, 1:3 positives to negatives)). Source: arxiv.org. Saturation forecast: Around February 2027. 4 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | DeepSeek V4 Pro | 75.2 | #181 |
| 2 | MiniMax-M3 | 75.2 | #265 |
| 3 | DeepSeek V4 Flash | 73.5 | #254 |
| 4 | Gemini 3.5 Flash Lite | 72.7 | #300 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=interpretive-canons-bverfg · How It Works · Data refreshed daily, snapshot 2026-11-12.