| # Program Execution Tracing | |
| Each instance contains 320 shuffled records of earlier Python computations and | |
| one final eight-step query. For every query step, the solver must find the | |
| semantically equivalent earlier function whose recorded call uses the required | |
| input. Earlier outputs route later steps. | |
| The structured prediction contains the final output under `answer` and the | |
| eight supporting cell IDs under `evidence`. Primary exact accuracy requires both | |
| the output and the exact supporting-cell set. | |