Model card: evaluation results (2026-09-29)

README.md

@@ -64,6 +64,21 @@ Temperatures do not change which outcome is most likely; they only make the prob
 - **Data:** an internal mixture of judgment tasks derived from public datasets together with internally generated
   and synthetic tasks. The training data is not released.
 
+## Evaluation
+
+Measured on 2026-09-29 on one NVIDIA RTX PRO 6000 (BF16), with the serving temperatures, one request at a time.
+
+| Benchmark | Protocol | Result |
+|---|---|---|
+| [JevBench](https://github.com/fstandhartinger/jevbench) v1.4, public items | 231 decisions: easy / standard / hard | 100% / 97.2% / 57.7% |
+| [DecisionBench](https://huggingface.co/datasets/Hanno-Labs/decision-bench) 1.0 | All 23,900 rows; the 251 rows that exceed the token budget count as wrong | 77.4% accuracy, ECE 0.060, NLL 0.72. On the 20,959 rows whose source text does not appear in our training data: 77.0% |
+| [typed-decisions](https://huggingface.co/datasets/LocalLLaMA/typed-decisions) | Test split, 400 cases / 2,000 decisions; zero-shot generalist (none of its workflows were trained) | Accuracy 0.594, soft accuracy 0.517, KL 0.288, Brier 0.143 |
+| [ImajevBench](https://huggingface.co/datasets/mohit67890/imajev-bench) v2.0-lite | Dev + calibration splits (254 items with public gold), direct option scoring with "unknown" appended, all cyclic option orders averaged | 66.9% (chance 26.5%); visual 76/103, text 24/38, joint 70/113 |
+| [behavior-benchmark](https://huggingface.co/datasets/respanai/behavior-benchmark) | Every trace, three-way choice (present / absent / not observable); F1 of `present` | Core 0.458, multilingual 0.668 |
+| Fair-chance items (18 kinds: coins, dice, lotteries, tie-breaks, …) | Mean KL from the uniform distribution | 0.0009 |
+
+Median latency on the JevBench public items is 0.039 s per decision (p95 0.123 s).
+
 ## Limitations
 
 - **Answer verification.** Asked whether a free-form worked answer (for example a maths solution ending in a