Datasets:
| # Repair episodes | |
| An episode links one versioned broken input to a model's actions and final submitted package. Multiple models and repeated attempts can share a task without overwriting each other's records. | |
| Record `case_id`, `task_revision`, `model_config_id`, `attempt_id` and `episode_id`. Retain the input and final package hashes, model identifier and reasoning settings, harness version and verifier identity. | |
| Turn 0 records the input score. For each model turn, keep the response, exact recorded tool results, delivered progress, candidate package hash, criterion contributions, repair reward and change from the preceding measured state. Include token usage and cost when available. Mark truncation in an actor-facing tool response separately from the full retained tool output. | |
| The final report records submission validity, preservation, unresolved requirements and the final partial-credit reward. Missing measurements remain null with a reason. An unchanged package may reuse its existing measurement, but a modified package must be checked again. | |
| These records describe the final asset and the route taken to repair it. The final score is never replaced by an earlier, better checkpoint. Numeric rewards are retained for analysis, while the repair agent receives the qualitative progress configured by its harness. | |
| The per-turn fields require the SimRepairBench repair-progress integration. A standard Harbor run can use the same tasks with terminal verification without producing this checkpoint history. | |