AI systems deployed over large codebases treat a storage parameter as a cognitive one. Context window size determines what a model processes — it does not determine what the model attends to reliably at the moment of a specific decision. This note names the structural distinction between context availability and comprehension capacity, documents the failure class their conflation produces across AI coding tools and long-running sessions, and proposes a mechanism. The conflation was architecturally invisible when context was small because a developer curation layer was performing the selection function implicitly. As context windows grew and that layer was abandoned, the assumption became productive: loading more context activates the failure condition while appearing to address it. A finite diagnostic procedure is documented in the restricted version of this record.
Roman Kir (Mon,) studied this question.