| # References: the attorneys' redlines |
|
|
| One Word file per task, named after the task: `references/<task name>.docx` is the redline that the task's attorney |
| produced for that task, as submitted and then anonymized (tracked-change and comment authors, document properties and |
| comment initials carry only the party label, such as "WhitHill Legal"). |
|
|
| What they are for: |
|
|
| - They are the human reference for each task. The benchmark does not grade a model against them: grading uses the |
| attorney's rubric items in `tasks/<task name>/tests/rubrics.json`, which the same attorney wrote after finishing the |
| redline. |
| - Nothing in a task directory reads them, and the task's Docker image never contains them. They sit outside `tasks/` |
| so that an agent running a task cannot see them. |
|
|
| How they relate to the tasks: |
|
|
| - A Turn 1 reference is the attorney's opening redline of the clean contract in `tasks/<task name>/environment/app/contract.docx`. |
| - A Turn 2 task starts from a counterparty redline. That input is the Turn 1 reference of the task it answers |
| (`responds_to` in the task's `task.toml`), byte for byte, except in the two-party relabelled scenario noted in the |
| dataset card. The Turn 2 reference is the attorney's response on top of it. |
| - A few references contain edits the attorney made with Track Changes switched off. They are kept as submitted; the |
| rubric items, not the redline, define what the task grades. |
|
|
