| Decision rows for the pointer task. | |
| These are labelled System One requests imported from a published suite: yes/no, choice, and score questions. Choice questions with more than 26 options are omitted, which drops the 77-way banking block. | |
| `decisions/train.jsonl`, `validation.jsonl`, and `calibration.jsonl` are the training splits. `decisions/test.jsonl` is the in-domain test. It is scored and not trained on. `transfer-dev.jsonl` is a held-out transfer development set whose contexts do not appear in train. | |
| Structured state, instructions, and criteria are flattened to labeled lines (`key: value`), the same rendering the service uses when it scores a request. | |