| name: Science Prediction Universe | |
| description: > | |
| Science Prediction Universe (SPU) tests whether AI agents can predict the outcomes of | |
| real wet-lab experiments. Each of the 100 public tasks asks the agent to rank | |
| experimental candidates by a measured outcome, and rankings are scored by | |
| Spearman correlation. Report the number of scored tasks, model version, | |
| reasoning effort and harness with every result. | |
| evaluation_framework: harbor | |
| tasks: | |
| - id: public100-scored-mean-spearman-v1 | |
| config: default | |
| split: test | |