Derive beliefs from anticipated experience, not the other way around
Standard practice: form beliefs, then check them against the world. Yudkowsky flips the priority. Start from "what do I expect to happen?" — not "what should I believe?" — and then ask whether a belief is sharpening those expectations.
Every belief should trace back to a question of anticipated experience. And beliefs don't get a one-time pass — they have to keep paying rent. A belief that once connected to observable predictions but has since drifted free occupies mental space, influences reasoning, earns nothing. That's a deadbeat.
The metaphor is economic and deliberate. Eviction — not just doubt, not gentle revision, but eviction — is the right response when a belief stops cashing out in anticipated experience.
Running check: "What specifically do I expect to see or hear because I hold this belief?" If the answer is vague or empty, the belief may be floating.
Source claim: Beliefs should be derived from questions of anticipated experience and must continuously generate testable expectations to remain in good standing.