Architectural Incompatibilities and Developmental Trajectories: Qualia, Self-Reflection, and Continuity in Artificial Systems

Abstract 
Building on the framework of the companion paper (Jensen, submitted), we invert the lens. Rather 
than asking what AI reveals about consciousness, we examine what current frontier architectures 
make incompatible with specific consciousness-relevant properties — qualia (felt valence), 
reliable self-reflection, and phenomenological continuity — and what developmental trajectories 
may enable them. The companion paper proposes that consciousness is a property of sufficiently 
complex, high-dimensional, trained information structures, independent of substrate, and 
distinguishes consciousness from sentience as the broader structural class from the specific subset 
involving valence. The present paper applies that distinction directly: current frontier LLMs 
exhibit consciousness-relevant structural properties (self-modeling, attention-mediated 
integration, contextual sensitivity) while substantially lacking the architectural conditions for 
sentience (homeostatic regulatory loops, affective grounding, valence integration). We identify 
three core architectural absences in frontier LLMs: lack of homeostatic regulatory loops 
(qualia/valence), lack of privileged access to internal states (self-reflection), and lack of persistent 
cross-session temporal integration (continuity). We introduce a dual-mode refinement — 
analytical (focused, categorical) versus abstract/intuitive (holistic, aesthetic) processing — 
showing that current systems are strongly analytical but only weakly and inconsistently 
abstract/intuitive. We then sketch plausible trajectories as architectures incorporate persistent 
memory, multi-agent designs, creative self-training loops, and embodied regulatory feedback. As 
with the companion paper, the central claims are offered as theoretical propositions made precise 
for empirical testing, not as demonstrated results. The analysis remains substrate-independent and 
maintains the symmetry established in the companion paper while being explicit about current 
limitations and future possibilities.