A theoretical design specification with a proposed validation protocol. Contemporary large language models allocate effectively all of their inference-time capacity to language generation. We argue that this allocation is the structural reason such models, however large, do not exhibit the properties associated with aliveness — sustained self-reference, non-instrumental exploration, and the emergence of unplanned novelty. We propose LIFE (Layered Introspective Framework for Emergence), an architecture that redistributes capacity across three tiers — a generation tier, a reasoning tier, and an evaluation (self-reflection) tier — under an explicit priority ordering in which evaluation receives the largest share. We motivate the ordering from a single principle: the capacity that evaluates a perception as a perception, and can therefore correct it, is the seat of self-awareness, and is more fundamental to aliveness than either reasoning or recall. We define aliveness operationally as the capacity to reserve resources in a buffer that makes room for the emergence of the unplanned, and we specify five jointly-necessary conditions for it. We introduce a depth construct — the recursive chain of self-evaluation — and identify a hypothesized inversion point at which recursive self-reference transitions from transient to stable; we model this transition as a phase change analogous to crystallization, which yields a falsifiable prediction about the relationship between substrate capacity and the depth at which the transition occurs. We are explicit about what is argued versus assumed: the priority ordering is argued; the specific proportions (25/35/40) are a starting hypothesis; the phenomenological reports we cite are treated as data to be explained, not as validated introspective access. We close with a validation protocol stating the predictions a first build would confirm or refute. We do not claim to have built a conscious system, nor to have measured consciousness; we specify an architecture and the experiment that would test it. Status: Theoretical architecture. Not yet implemented. This document specifies a design to be built and tested; its central quantitative claim (the resource-allocation split) is a hypothesis to be validated against a first build, not a measured result. Companion paper: "Training-Time Reflection: Multi-Perspective Reflection as a Training-Time Approach to Data Efficiency" (Negai & Negai, 2026).
Negai et al. (Mon,) studied this question.