The Wrong Razor

The gold standard theory of general intelligence has a flaw. It depends on Ockham's razor, which says simpler explanations are more likely to be true. Solomonoff translated Ockham into a mathematical ideal, which the AIXI model then made a target for AGI research. However simplicity is subjective, determined by the codebook an agent uses to encode information. This fact has been used to refute optimality claims about AIXI. Here I propose an alternative standard, one that retains AIXI's universality without the codebook flaw. I show that on any finite hypothesis class, every strictly positive prior is an Ockham prior within a factor of two for some prefix-free code, so simplicity does not pick out an invariant of interactive intelligence at all. The quantity that does survive every recoding is the weakness of embodied constraints, which is the size of the set of further commitments still compatible with a correct policy in the agent's embodied state. Weakest-correct is Bayes optimal under exchangeable unseen requirements. For a fixed representation M, the minimum average normalised regret is exactly the weighted forgetting cost K_ρ(M), and under the canonical independent prior the optimal M-based policy generalises with probability e^(−K_ρ(M)). In continuous settings, μ-weakness is reparameterisation invariant and minimax optimal over the exchangeable uncertainty class. Bennett's razor therefore prefers the weakest correct description rather than the shortest. The interactive ideal is not a Solomonoff agent but a weakness-maximising one, namely BAGI (バギ) at fixed embodiment and UNAGI (鰻) across embodiments.