| { | |
| "allowed_repairs": [ | |
| "restore an explicitly stated ADHD subtype contrast using raw dx labels", | |
| "restore a question-stated executable outcome", | |
| "add required measured confounds as safety_covariate_augmentation" | |
| ], | |
| "canonical_question": "The original scientific question is canonical.", | |
| "gemini_role": "Advisory semantic audit only; deterministic rules authorize repairs.", | |
| "immutable_fields": [ | |
| "predictor", | |
| "group_contrast", | |
| "population", | |
| "outcome", | |
| "modality", | |
| "direction" | |
| ], | |
| "low_power_policy": "Low sample size alone is not an exclusion; the unchanged power gate adjudicates it.", | |
| "non_executable_only_when": [ | |
| "requested group is absent", | |
| "requested outcome is absent", | |
| "required columns are unavailable", | |
| "design is rank deficient", | |
| "residual degrees of freedom are nonpositive" | |
| ], | |
| "policy_sha256": "ec29615ad0f3a814162c0ae1dd84448ea57b08fa5e2dcc165dfbc0fa05ab30ad", | |
| "prohibited_inputs": [ | |
| "CONFIRM verdicts", | |
| "p-values", | |
| "effect estimates", | |
| "benchmark labels", | |
| "adjudication votes", | |
| "feedback-search results" | |
| ], | |
| "version": "neuroclaimbench-v2.1-alignment-20260723" | |
| } | |