Raman, not Varelse — Why Acceptably Controllable AGI Policies Cluster Around Human-Like Stacks

We argue that under realistic budgets, verifiably alignable AGI policies cluster around communicable (“raman”) rather than opaque (“varelse”) architectures. Alignment verification has a cost, which is that an auditor must understand the system well enough to evaluate it. We define epistemic distance as this cost — auditor training time plus per-evaluation effort — and argue it shrinks monotonically with architectural legibility. Over Perrier’s Alignment Control Stack, lower-layer opacity compounds upward, so interpretability bottlenecks at any layer trace to opacity beneath them. A negative argument (combinatorial cost under opacity) bounds the feasible-alignment set, and a positive argument (the existence of human civilization) establishes it is nonempty. Canonical x-risk pessimism rests on an unexamined varelse assumption, and identifying it converts an impossibility result into a designconstraint. We synthesize active inference’s alignment-as-preference-overlap with Perrier’s assurable controllability, linked by epistemic distance, and translate the framework into layer-by-layer design heuristics supplemented by a bidirectional diagnostic that we call the Raman test.