It’s straightforward to compute a single vector of all the a * P["innovation"] values in one go—I don’t then know whether you could unfold the Dirichlet into something vectorized in PyMC.
a = [P["start"] vi]
This has the same number of degrees of freedom (N - 1 from simplex, from concentration) and is usually a much better-behaved parameterization of Dirichlet (and the Beta, which is just a two-element Dirichlet that only returns one of the elements). There’s a discussion of this in Chapter 5 of Gelman et al.'s Bayesian Data Analysis around the first hierarchical model introduced there for rat clinical trials, where they reparameterize as a mean and concentration and discusses priors.