How a Predictive State Observer Can Self-Adapt Its Sensory Prediction-Error Correction Gain: Closed-Loop Evidence from a Muscle-Driven Reaching Task

We ask how a forward-model-based predictive state observer should set its sensory prediction-error correction gain during muscle-driven reaching, and whether that gain can be adapted from agent-available signals -- innovation history and per-episode reaching outcome -- rather than from swept oracle labels. We evaluate a residual-MLP forward model in a 34-muscle MyoSuite arm on an IK-reachable below-shoulder task, deployed in closed loop with a stabilized endpoint probe controller that uses non-negative least-squares muscle routing and a virtual target ramp; the controller is a stabilized probe for evaluating state-estimation effects, not a biological motor planner. A swept fixed-gain closed-loop oracle reveals a delay-dependent correction structure: with no sensory delay, intermediate correction gains are best (K = 0.25-0.50), whereas with 18-step delay observation-heavy correction wins (K = 1.0). The forward-model-only K = 0 ablation is not the oracle: it is systematically worse than the best fixed K by 1.9-6.1 cm and shows large NNLS controller residuals caused by long-horizon autoregressive drift; we therefore report K = 0 as a diagnostic. Outcome-trained reliability-adaptive observers improve the delayed regime by 1.9-2.5 cm over default reliability while remaining neutral in no-delay cells, where the oracle is already intermediate. A feature-conditioned {beta} adapter that maps cell-level innovation statistics to per-field gain parameters nearly matches a per-cell trained diagnostic in 5/6 cells, but both remain 1.4-1.8 cm worse than the swept fixed-K oracle at 18-step delay. These results separate the delay-dependent correction structure, the forward-model-only failure mode of K = 0, and the remaining limits of agent-available adaptive correction.