Eliciting Latent Knowledge Via Hypothetical Sensors — LessWrong