Four samples of Fe-bearing prismatic sillimanite, containing ∼1 wt.% Fe2O3, were annealed experimentally at temperatures of 1465 and 1675°C, and pressures between 1 atm and 30 kbar. Transmission electron microscopy (TEM) and analytical TEM (ATEM) investigation of the samples reveal that the starting material partly transformed into mullite during the annealing, and that this process was assisted by partial melting. The exsolved partial melt (now a glass), observed at triple junctions and in the form of small precipitates (∼10–1000 nm in size) within the sillimanite matrix, contains >80 wt.% SiO2. It also contains ∼11 wt.% Al2O3, some FeO and detectable amounts of K2O and CaO. Dissociated c dislocations in sillimanite are preferential nucleation sites for SiO2-rich precipitates. The equilibrium compositions of residual sillimanite-mullite were measured with a 2 nm wide probe at the interface with the SiO2-rich glass in each sample after heat treatment. We used these equilibrium compositions to constrain the parameters of a point defect model for sillimanite mullitization proposed by Raterron et al. (1999). With the revised parameterization, it is now possible to calculate the position of the boundary between fields of mullite + melt and mullite in the SiO2–Al2O3 phase diagram, and to predict the effect of pressure on this boundary. However, to be used as a standard, this model still needs to be calibrated in the pure SiO2–Al2O3 system (without impurities such as iron).