Causal Understanding as Domain-Relative Interface Problem. Why Input Channels, Not Architecture, Limit AI Reasoning.

This paper revises my earlier thesis that Transformer architectures are structurally incapable of causal reasoning. I now argue that causal understanding is domain-relative and interface-limited, not architecturally constrained. An AI system's ability to reason causally in a domain scales with its interface completeness—whether it has access to intervention→consequence feedback, not merely descriptions. This framework yields three novel claims: (1) the domain-relativity thesis—causal competence is not a global property but varies by domain according to interface richness; (2) the programming paradigm—code is "easy" for AI because it is a fully text-embedded causal domain where the intervention→consequence loop closes entirely in tokens; (3) the difficulty inversion principle—AI and human difficulty orderings are roughly inverse because difficulty scales with interface completeness for machines but abstraction level for humans. I derive testable predictions and connect the framework to world models. From my original thesis, I preserve the distinction between operational causal models and felt causal relevance—the latter still requires interoceptive embodiment.