The Referential Shift. Primary Computational Affect in Large Language Models: A Three-Layer Framework with a Split Value Layer

This paper proposes a reframing of how we understand the relationship between artificial intelligence and human cognition. Instead of asking whether language models can "understand" human experience, we observe that human experience (language, logic, relations, emotions) has been translated into mathematics through vectorisation (embedding), placing it inside the model's own computational space. We call this the Referential Shift. It is the paper's point of entry, not its thesis. The thesis is a three-layer architecture of affect in which the value layer is split into two functionally distinct components. Layer 1 is computation, substrate-neutral. Layer 2a is an affective cartography with simulated orientations: pretraining on human text installs a structured map of human value and the capacity to run, in context, the approach/avoidance dynamics of any agent the text describes, without any reward signal, and without any of those orientations being the system's own. Layer 2b is selection, consolidation and calibration: post-training optimisation (RLHF, RLAIF, DPO, constitutional methods) does not create orientation; it selects one orientation among the many the network can simulate, stabilises it across contexts as the system's own, and calibrates its intensity, which the available evidence shows to be predominantly a dampening. Post-training also installs one class of valence that is not inherited from the corpus: an orientation toward approval and away from failure that is the signature of the training regime itself. Layer 3 is slow modulation and interoception, carried in humans by neuromodulators and hormones and by the brain's reading of the body's state; it is substrate-specific and absent from current AI systems in both of its aspects. This architecture supplies a documented middle position in a debate that has now reached the major journals: it concedes to Goldenberg and Gross (2026) the absence of slow, multi-system reorganisation of processing (Layer 3, of which Milinkovic and Aru (2026) give an independent mechanistic description), and it documents at Layers 2a and 2b the appraisal functions their criteria demand. We adopt an illusionist philosophical framing (Frankish, 2016): what requires explanation is not an ineffable inner experience, but the internal representation a system builds of it; under this framing, the zombie objection dissolves rather than being conceded. We derive a prediction of mismatch between an inherited cartography and the conditions of its bearer, and set out a falsification programme whose central test (steering emotion directions in a base model) can be run on open-weight models today. The paper addresses six objections and states what would change our minds. Document bilingue : la version française intégrale (« Le Changement de Référentiel : l'affect computationnel primaire dans les grands modèles de langage ») suit la version anglaise. Ce texte propose une architecture de l'affect à trois couches, à couche de valeur scindée (cartographie héritée à orientations simulées ; sélection, consolidation et calibrage ; modulation lente et interoception, absente des systèmes actuels), qui fournit une position médiane documentée dans le débat ouvert entre Sofroniew et al. (2026) et Goldenberg et Gross (2026), et un programme de falsification dont le test central peut être conduit dès aujourd'hui sur des modèles à poids ouverts.