| # Task data | |
| Each benchmark task gets one TOML file under `manifests/`. The example is a | |
| schema template, not an eligible experiment task. Split files contain one | |
| `task_id` per line and must be frozen before confirmatory evaluation. | |
| Gold patches, gold locations, and hidden tests are evaluator-only data. The | |
| agent receives only the problem statement and repository state, except in the | |
| explicit oracle controls. Public repository URLs and exact base commits are | |
| required for reproducibility; obtain and preserve any repository license and | |
| redistribution metadata before publishing a task bundle. | |