Co-folding models execute covalent bonds they do not propose, and pose RMSD certifies the difference away

Co-folding models are evaluated on ligand pose RMSD. For a covalent complex that metric does not measure the bond, and the gap between the two widens as pose prediction improves. Across 21 KRAS entries carrying a deposited protein-ligand covalent bond and released after the Boltz-2 training cutoff, with the bond not declared, Chai-1 reaches covalent bond distance in 60 of 75 cysteine samples and 0 of 30 on Tyr71 and Asp12; Boltz-2 in 9 of 75 and 0 of 30; Protenix v1.0.0 in 4 of 75 and 0 of 30. Declaring the bond recovers Chai-1 to 75/75 on cysteine and 12/15 on Tyr71, and Protenix to 75/75 and 15/15, so the restraint machinery works and the gap lies in what the models propose unprompted. No model proposes a covalent bond off cysteine; all three execute one when told; and on 15 of 21 ligands one of the three cannot be told at all, because Boltz-2's bond constraint requires a CCD entry and its dictionary is frozen at build time. The gap is not training-set leakage. On a matched arm of 40 pre-cutoff cysteine entries, no model scores meaningfully higher on the data it trained on - Chai-1 57.0% against 57.1% post-cutoff, Boltz-2 6.0% against 11.4%, Protenix v1.0.0 7.0% against 0%, the last inside what 35 held-out samples can resolve - and two Protenix checkpoints on one architecture, straddling a tier boundary, score higher on the tier still held out for the later one, which is the ordering memorization would reverse. Boltz-2's median on those controls is 2.31 A, so on structures released up to ten years before its cutoff it still stops at contact distance without bonding: the learned non-bonded stop is not an artifact of ligand novelty. On 360 samples with symmetry-corrected pose RMSD computed alongside bond distance, 45 are accepted by a 2 A pose cut and rejected by the bond metric, and on 9PZY the declared and undeclared arms are pose-indistinguishable (0.378-0.572 A against 0.422-0.597 A) while completely bond-separated (1.627-1.669 A against 2.684-2.813 A). The better pose predictor of the two Protenix checkpoints is where the disagreement is largest. Improving pose accuracy does not improve bond accuracy; it raises the rate at which the pose metric certifies structures that have no bond. Two large 2026 benchmarks of covalent co-folding, covering 2,390 complexes between them, report pose metrics only, and both reach conclusions that do not survive measuring the bond. The gap also cannot be closed downstream: the atom naming needed to declare a bond differs between every tool and the deposited file, and in the six-entry Runs N' Poses set the bond record a covalent refinement step would need is absent from all six files and was never recorded in five. Files: the manuscript (PDF), and supplementary data, code and analysis notes (ZIP) - per-sample results for all four model checkpoints, both arms, including the 40-entry control arm for three of them and pose RMSD on 360 Protenix samples; run manifests; the symmetry-corrected pose RMSD implementation; the four atom-naming schemes reimplemented from installed source; and nine documented silent failure modes.