{"schema_version":1,"experiment_id":"eoa_inverse_innovation_exp04_retrieval_first_paired20_20260802","cell_id":"invariant_mode_decomposition_design__psychology","considered_ids":["H1","H2"],"selected_id":"H2","dispositions":[{"hypothesis_id":"H1","lane":"COLLISION","reason":"Substantial collision with existing person-specific EMA control theory, eigenvalue/eigenvector targeting, and idiographic CBT-module selection. Its residual conjunction is bounded-search-only, problem support is moderate, and the claimed 0.3-standardized-unit advantage lacks direct supporting evidence in the critic record."},{"hypothesis_id":"H2","lane":"FINALIST","reason":"Despite substantial collision with adaptive exposure titration and physiology-driven controllers, the critic identifies a material, falsifiable residual: jointly estimating trial-level fear, approach, avoidance, arousal, and safety-behavior dynamics, classifying invariant modes by stability and intervention sensitivity, and using those classifications to advance, hold, or regress exposure difficulty. Direct sources also establish a strong, recognizable pacing problem."}],"selection_reason":"H2 offers the narrower and better-supported differentiated opportunity. Its residual decision rule is explicit and falsifiable, while the cited literature directly supports both the crowded neighboring practice and the clinical inadequacy of scalar fear or anxiety signals alone. H1's remaining distinction is similarly collision-heavy but rests on only moderate problem support and adds an unsupported effect-size claim."}