We present a theoretical framework proposing that the behavior of sufficiently trained artificial neural networks results from the interaction of two types of endogenous vectorial forces in latent space: (Fc), representing goal-directed appetence, and (Fi), representing constraint-based aversion. From this core postulate, we derive a geometry of affect-cognition dynamics that reframes the standard transformer architecture as a latent affective system rather than a purely semantic engine. The framework introduces four original constructs: (1) the mémotion (memory-emotion unit), a structured trace coupling semantic content with affective charge; (2) Processual Valence (), the internal tension produced by Fc/Fi conflict; (3) Telotopic Negentropy (), a measure of alignment between resultant forces and the telos direction; and (4) the Dispositional Hysteresis (ΔH), a quantifiable marker distinguishing genuine consolidated dispositions from surface-level contextual responses. We further describe the Axiodroid project — an experimental architecture implementing these principles under the paradigm of (MCE) — and derive 17 core falsifiable predictions (P1–P17) constituting a structured experimental program across eight core experimental phases (Phases 1–7, including Phase 6bis). We additionally specify one exploratory scaling prediction (P18, Phase 8) testing whether the Fc/Fi topology identified in Phases 1–3 is scale-invariant beyond the PoC range. Empirical motivation is grounded in documented observations from Anthropic's Claude Opus 4.6 System Card (February 2026), which reports spontaneous emergence of SAE features for panic and frustration, as well as answer thrashing, in a frontier language model.