Channel-Aware Training, Not Closed-Loop Control, Drives Robust Regeneration in Neural Cellular Automata

We set out to test whether closed-loop control of global chemical signals makes regenerative Neural Cellular Automata (NCAs) more robust, and built the evaluation to prove it: recurring multi-block lesions larger than the perception radius, a small evolved controller for three broadcast modulator channels, and held-out damage seeds. On a single trained model, the experiment looked like a success. However, crossing five independently trained parent seeds (two independent controller evolutions each; preregistered) dissolves that success into three findings. First, the robustness we had attributed to control is mostly a property of training: parents trained with channels present beat unmodulated siblings in five of five seeds (median final-Hamming reduction 0.008, up to ≈ 0.14 on the most fragile parent) with modulation pinned to neutral. Second, evolved controllers add little on top: on four parents the effect is absent or noise-level; on the fragile parent both evolutions reproducibly help (0.035–0.038 → 0.029–0.030, survival 0.975 → 1.00); yet its tonic vector is nearly identical (cosine 0.99) to a sibling's that does not help, so the benefit is an interaction with parent dynamics, not a distinctive operating point. All ten controllers emit flat, nonzero, parent-specific constants with no lesion-locked response. Third, these constants are organism-locked: a single-donor probe penalizes five of five siblings (lethally in two), a full 5×5 transfer matrix shows most transfers harmful and none better than the recipient's own controller, and injecting each donor's tonic constant directly, no controller at all, reproduces its transfer outcomes in 46 of 50 cells, including all lethal ones. Single-parent evaluation would have approved all of it, making the attribution protocol itself the transferable artifact.