Comprehension and Production

Transformer language model inference consists of two architecturally distinct phases: prefill (parallel processing of the input prompt under causal masking) and generation (sequential, autoregressive production of output tokens). These phases differ in their computational character, parallel versus sequential, comprehension versus production, all-at-once versus one-at-a-time, yet they are executed by the same model weights and have never been systematically compared at the activation level. This paper presents a pre-registered experimental design for characterising the meta-dimensional dynamics of each phase, the transition between them, and how the transition's character varies across prompt types. Using the Cognitive Weather Observatory (Temte 20260), the experiment tests whether prefill and generation constitute observably distinct processing regimes, whether the transition between them exhibits a characteristic phase-transition signature, and whether the nature of the prompt modulates the transition's character. The design addresses a question both technically foundational and philosophically suggestive: does the system process differently when it is understanding versus when it is producing, and if so, what does the boundary between these modes look like?