Since the availability of assembled eukaryotic genomes, the first one being a budding yeast, many computational methods for the reconstruction of ancestral karyotypes and gene orders have been developed. The difficulty has always been to assess their reliability, since we often miss a good knowledge of the true ancestral genomes to compare their results to, as well as a good knowledge of the evolutionary mechanisms to test them on realistic simulated data. In this study, we propose some measures of reliability of several kinds of methods, and apply them to infer and analyse the architectures of two ancestral yeast genomes, based on the sequence of seven assembled extant ones. The pre-duplication common ancestor of S. cerevisiae and C. glabrata has been inferred manually by Gordon et al. (Plos Genet. 2009). We show why, in this case, a good convergence of the methods is explained by some properties of the data, and why results are reliable. In another study, Jean et al. (J. Comput Biol. 2009) proposed an ancestral architecture of the last common ancestor of S. kluyveri, K. thermotolerans, K. lactis, A. gossypii, and Z. rouxii inferred by a computational method. In this case, we show that the dataset does not seem to contain enough information to infer a reliable architecture, and we construct a higher resolution dataset which gives a good reliability on a new ancestral configuration.