A Safety-Governed Architecture for Adaptive Sequential Evidence Selection from Pre-Recorded Gait Data

Sequential evidence selection can reveal which parts of an existing sensor record reduce model uncertainty, while providing a computational test bed for future adaptive sensing. This work introduces the Embodied Evidence Acquisition and Reasoning Loop (EARL), a typed architecture in which five role-specialized critics score evidence requests and a deterministic safety governor retains exclusive execution authority. Observed, derived, and simulated evidence remain provenance-distinct in a replayable, hash-linked ledger. The primary experiment sequentially disclosed precomputed feature bundles from pre-recorded gait data; it did not acquire new measurements. EARL was evaluated on 64 unique subjects from the PhysioNet Gait in Neurodegenerative Disease Database using repeated subject-level cross-validation, 11 predeclared conditions, and 3520 replay-verified runs. The primary endpoint was area under cumulative posterior-entropy reduction. EARL achieved 7.689 (95% CI 7.535-7.840), exceeding fixed-order and random selection by 0.459 and 0.763, respectively; the EARL-minus-EIG difference was -0.321. EARL had descriptively higher macro accuracy (55.8% versus 49.5%) and a lower Brier score (0.781 versus 0.813) than pure expected-information-gain selection. This measures an entropy-efficiency trade-off under additional selection criteria, not universal superiority. All 10 original software safety challenges produced their expected outcomes. Separately identified post hoc analyses examine probe use, a sensitivity analysis excluding the record-quality probe, critic influence, illustrative resource costs, early stopping, and safety-constrained simulated execution. The small retrospective cohort and illustrative simulations establish software behavior, not clinical diagnostic performance, treatment benefit, physical-robot safety, or patient efficacy.