<jats:title>Abstract</jats:title><jats:p>The recent publication of an updated thermodynamic dataset for petrological calculations provides an opportunity to illustrate the relationship between experimental data and the dataset, in the context of a new set of activity–composition models for several key minerals. These models represent orthopyroxene, clinopyroxene and garnet in the system CaO–MgO–Al<jats:sub>2</jats:sub>O<jats:sub>3</jats:sub>–SiO<jats:sub>2</jats:sub> (CMAS), and are valid up to 50 kbar and at least 1800 °C; they are the first high‐temperature models for these phases to be developed for the Holland & Powell dataset. The models are calibrated with reference to phase‐relation data in the subsystems CaO–MgO–SiO<jats:sub>2</jats:sub> (CMS) and MgO–Al<jats:sub>2</jats:sub>O<jats:sub>3</jats:sub>–SiO<jats:sub>2</jats:sub> (MAS), and will themselves form the basis of models in larger systems, suitable for calculating phase equilibria in the crust and mantle. In the course of calibrating the models, it was necessary to consider the reaction orthopyroxene + clinopyroxene + spinel = garnet + forsterite in CMAS, representing a univariant transition between simple spinel and garnet lherzolite assemblages. The high‐temperature segment of this reaction has been much disputed. We offer a powerful thermodynamic argument relating this reaction to the equivalent reaction in MAS, that forces us to choose between good model fits to the data in MAS or to the more recent data in CMAS. We favour the fit to the MAS data, preserving conformity with a large body of experimental and thermodynamic data that are incorporated as constraints on the activity–composition modelling via the internally consistent thermodynamic dataset.</jats:p>