The recently proposed ATOMIC protocol is a fully ab initio thermochemical approach designed to provide accurate atomization energies for molecules with well-defined valence structures. It makes consistent use of the concept of bond-separation reactions to supply high-level precomputed bond increments which correct for errors of lower-level models. The present work extends the approach to the calculation of standard heats of formation and validates it by comparison to experimental and benchmark level ab initio data reported in the literature. Standard heats of formation ((2) over bar 98 K) have been compiled for a large sample of 173 neutral molecules containing hydrogen and first-row atoms (C, N, O, F), resorting to several 'previous compilations and to the original experimental literature. Statistical evaluation shows that the simplest implementation of the ATOMIC protocol (composite model Q achieves an accuracy comparable to the popular Gaussian-3 approach and that composite models A and B perform better. Chemical accuracy (1-2 kcal/mol) is normally achieved even for larger systems with about 10 non-hydrogen atoms and for systems with charge-separated valence structures, bearing testimony to the robustness of the bond-separation reaction model. Effects of conformational averaging have been examined in detail for the series of n-alkanes, and our most refined composite model A reproduces experimental heats of formation quantitatively, provided that conformational averaging is properly accounted for. Several cases of larger discrepancy with respect to experiment are discussed, and potential weaknesses of the approach are identified.