Reference level of Y for multinomial regression

I’m fitting a multinomial logit mode to an outcome with 4 categories, and I’m estimating K-1 intercepts. I’m also specifying p[K]=0 to restrict the 4th category to a fixed value. The outcome categories are labeled 1-4 in stan data and it looks like Stan uses the last K=4 category as a reference. I get three coefficients for alpha - a1, a2, a3. How are these ordered - for example is a1 an intercept for K=1, or is a1 the next closest level to the reference, which would be K=3?

Related, is there any reference material on how Stan handles categorical outcomes/reference values and how/if one can modify that?

I think I see it, it’s literally the numeric ordering. So looks like largest category is always the reference, and then a1 is for category 1, a2 - for category 2 and so on.

Also whenever there are say two predictors - beta[2,1] for example means beta for the second predictor and 1st outcome category.