# Supplementary Figure S14

**Repeated balanced-fold stability of the primary safe model.** Twenty deterministic five-fold realizations (one frozen primary and 19 prespecified sensitivity seeds) each assign all 2,160 targets once out of fold. Points show the realization-specific mean paired effect, horizontal coloured lines show the mean across realizations, shaded bands show joint hierarchical-bootstrap 95% intervals incorporating target- and realization-level uncertainty, and dashed lines mark zero. **(a)** RMSE difference versus zero and **(b)** RMSE difference versus train mean remain below zero. **(c)** All 20 raw-cosine point estimates are positive, but the joint interval (-0.000216 to 0.002498) crosses zero; this endpoint is therefore mixed rather than fold-robust directional evidence. **(d)** Perturbation-specific residual cosine remains above zero. The diamond denotes the frozen primary realization.