The Wrong Razor

Occam's razor says prefer the shortest correct description. The interactive Solomonoff family turned that razor into a theorem. AIXI and its embedded descendants used that to formalise the gold standard model of superintelligence. But the theorem smuggled in a codebook. On finite hypothesis classes every strictly positive prior lies within a factor of two of an Occam prior for some prefix-free code. Without a privileged abstraction layer, intrinsic semantic length ceases to discriminate, so any equal-score-symmetric selector based only on it collapses to uniform. Description length therefore does not identify the invariant of interactive intelligence. The standard defence is that invariance up to constants plus a natural reference machine. But that still leaves interactive policy ranking machine-dependent. This paper proposes a replacement. Bennett's razor says prefer the weakest correct description. Weakest-correct is Bayes optimal under exchangeable unseen-requirement growth. For a fixed representation M, minimum average normalised regret is K_ρ(M), and under the canonical independent relevance 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. The right interactive ideal is not a Solomonoff agent but a weakness-maximising one, namely BAGI (バギ) at fixed embodiment and UNAGI (鰻) across embodiments.