Abstract
<jats:p>Battery models can be degraded by EIS-derived parameters that fit spectra well but are not uniquely constrained: across 87 measured spectra from 13 commercial cells, equal-residual fits disagree on diffusion by 1.6–3.3 decades. We convert this hazard into risk-controlled parameter hand-off and measurement allocation: a coordinate-wise router selects each of 48 coordinates among eight regularized estimators, and an empirical risk controller accepts individual claims under a false-claim budget, behind out-of-distribution and frequency-support vetoes. Routing lifts development-panel recovery to 28.0 of 48 and replicates at 27.78, 26.82, and 27.82 on three disjoint, truth-isolated model-consistent panels generated after hash-freezing; the best validated point yields 10.38 correct claims per run at 0.960 realized precision. In a locked within-class transfer study, hand-off cuts median unseen-condition error from 0.163 to 0.018 versus documented defaults, yet the best predictor is a single prior-anchored fit recovering far fewer coordinates than routing: prediction accuracy does not certify parameters. A balanced 7 × 7 temperature–state-of-charge design matches full-frequency enrichment within the prespecified equivalence margin using 36% of the impedance points. On the measured cells the pipeline returns an auditable conditional record per cell while vetoing out-of-domain certification; real-cell parameter truth and hardware battery-management benefit remain the decisive open experiments.</jats:p>