Most phylogenetic models of protein evolution assume that sites are independent and identically distributed. Interactions between sites are ignored, and the likelihood can be conveniently calculated as the product of the individual site likelihoods. The calculation considers all possible transition paths (also called substitution histories or mappings) that are consistent with the observed states at the terminals, and the probability density of any particular reconstruction depends on the substitution model. The likelihood is the integral of the probability density of each substitution history taken over all possible histories that are consistent with the observed data. We investigated the extent to which transition paths that are incompatible with a protein's three-dimensional structure contribute to the likelihood. Several empirical amino acid models were tested for sequence pairs of different degrees of divergence. When simulating substitutional histories starting from a real sequence, the structural integrity of the simulated sequences quickly disintegrated. This result indicates that simple models are clearly unable to capture the constraints on sequence evolution. However, when we sampled transition paths between real sequences from the posterior probability distribution according to these same models, we found that the sampled histories were largely consistent with the tertiary structure. This suggests that simple empirical substitution models may be adequate for interpolating changes between observed sequences during phylogenetic inference despite the fact that the models cannot predict the effects of structural constraints from first principles. This study is significant because it provides a quantitative assessment of the biological realism of substitution models from the perspective of protein structure, and it provides insight on the prospects for improving models of protein sequence evolution.