| schema_version = 1 | |
| experiment_id = "E16" | |
| name = "heldout_harness_rank_validation" | |
| mode = "protocol_interface" | |
| description = "Study 5 held-out validation of six deterministically selected harness archetypes on 17 outcome-blind fresh tasks." | |
| harness_ids = ["H000", "H001", "H002", "H007", "H008", "H014"] | |
| edit_interface_ids = ["P002", "P003"] | |
| repository_ids = ["R002", "R003"] | |
| model_ids = ["M002", "M003", "M004"] | |
| embedding_id = "EMB002" | |
| task_split = "study5_fresh" | |
| context_budgets = [65536] | |
| seeds = [0] | |
| repetitions = 1 | |
| max_tool_calls = 12 | |
| max_test_runs = 2 | |
| timeout_seconds = 1800 | |