We compared the models NS, FIM, and SST (Figure 1) in the four species by means of an ABC-RF algorithm and estimated demographic parameters for the most supported model. After checking for the evolution of the out-of-bag error of the RF, model selection and parameter estimation were computed using respectively 500 and 1,000 trees in each species. We found that NS had the higher posterior probability (p=0.84) for G. cuvier (Tables 1 and S7). In contrast, demographic histories of the three other species were best described by SST, with a posterior probability ranging from 0.53 to 0.88 (Tables 1 and S7). The estimated median number of migrants per generation Nm was 1.8 (95% CI: 0.7-3.0) for C. melanopterus, 6.6 (95% CI: 1.5-15.4) for C. limbatus, and 11.5 (95% CI: 3.0-22.0) for C. amblyrhynchos (Figure 2, Table 1). The posterior distribution of Nm strongly differed from the prior distribution and showed a clear unimodal peak with small credible intervals, and low mean square error (SME) and mean root square error (SMRE) in all three species (Figure 2, Table S8, suggesting that these estimates are highly reliable. Conversely, both TCOL and NANC had larger SME and SMRE errors in all species (Table S8), but it was only in C. melanopterus where posterior and prior distribution could not be distinguished (Figure 2). TCOL has a clear unimodal distribution in C. amblyrhynchos but a more disperse one (and with wider credible intervals) in C. limbatus (Figure 2, Table 1).