Inference-Time Projection for Physically Valid Biomolecular Diffusion Models

AlphaFold-3-style cofolding models achieve high structural accuracy but still produce physically invalid structures with steric clashes, distorted ligand geometry, stereochemical errors, or overlapping chains. We formulate physical validity as a constrained inference problem and introduce a lightweight projection operator applied to the denoised clean-coordinate estimate, $\hat{x}_0$. At each late reverse step the operator selects the most severely clashing atom pairs and applies bounded, symmetric push-apart corrections, leaving the learned denoiser and its weights untouched. Across five benchmarks (CASP15, CASP16, the PoseBusters monomer and complex sets, and the Boltz physical-validity test set) our projection raises the physical validity of Boltz-2 from $52$--$95\%$ to $97$--$100\%$, while every structural-accuracy metric remains statistically indistinguishable from the unmodified base model. These gains are achieved with minimal runtime and memory overhead, in contrast to approaches that require additional finetuning or costly steering procedures whose memory requirements can make cofolding large complexes impractical.