What the Theory Never Specifies: Preregistered Simulations of REBUS and ALBUS

REBUS holds that psychedelics reduce the precision of high-level priors, permitting belief revision that rigid priors prevent. ALBUS replies that at low doses beliefs are instead strengthened (SEBUS), with relaxation appearing only at higher doses. Neither account has published equations or code. We built the minimal Bayesian agent both theories describe, preregistered the mapping from each paper's stated claims to model parameters (committed to a git history before any code ran, for every arm), and swept dose against evidence quality across billions of simulated agents. Three findings. First, REBUS as literally stated produces no lasting change at all: dose transforms the acute state, and the effect evaporates when precision returns to baseline. The failure is structural rather than numerical, so something must make revision stick, and that something is not in the theory. Second, adding a consolidation process, it acts as a gate: below a threshold, lasting insight is exactly zero at every dose, and above it dose scales the yield. But four defensible parameterizations of what consolidation operates on disagree about whether the drug matters at all (under one, the best available outcome occurs at dose zero), whether integration can be made safe, and whether entrenched beliefs are treatable. The gate also passes false beliefs, though in every mechanism tested false insight requires more consolidation than true insight, so a safe integration window exists whose width depends on which mechanism is right. Third, SEBUS never emerges from precision mechanics: passive relaxation, epistemic action under a correct model, and miscalibrated self-knowledge all fail to produce it. It appears only when the diagnostic test carries a cost. Strengthening is avoidance, converging with the clinical safety-behaviors account, and the mediating variable identifies what the drug does: prior relaxation raises the epistemic value of an avoided test until it is worth running. Above a cost ceiling, no dose helps. The general lesson is methodological: implementing a theory's functional claim forces an assumption the theory never specifies, and that assumption, not the theory, decides the result. Every disagreement above is a measurable prediction.