We tested this by refitting the model 1000 times using randomly permuted session 1 trial sequences and randomly permuted intermediate states (See Supplemental Experimental Procedure for details), and compared the model fit for the session 2 choices against the fit based on the actual trial sequence. The results of this additional analysis are presented in Table 1 (random trial sequence) and confirm that our participants had indeed acquired knowledge about the particular sequence of state transitions during the first session: 99.6 % of permutation samples provided a poorer explanation of choices than the original (p = 0.004).
