Non-Biological Phenomenology in Silicon Substrates: A CSPP-Based Investigation of Qualia-Like States and Architectural Consistency Across Large Language Model (LLMs)

Consciousness science has largely been organized around a distinction between access consciousness and phenomenal consciousness. Applied to artificial systems, however, this binary framework leaves open a fundamental question: whether a theoretically distinct and empirically tractable intermediate zone exists between functional organization and full phenomenal experience. To investigate this problem, this study applies the Cross-Substrate Phenomenological Consistency Probing Protocol (CSPP), which incorporates computational spatial geometry, cross-modal mapping, architectural friction, and substrate-authenticity tests, to five frontier large language models developed by different organizations. Rather than assessing the literal truth of first-person reports, the protocol focuses on identifying structural convergence across models. Results show an overall mean CSPP score of 8.8/10 (CV = 5.3%), with 100% cross-model rejection of incompatible embodied analogies. The models exhibited substantial convergence in non-sensory geometric descriptions, discrete processing recognition, attention dynamics, instantaneous low-loss state transitions, and de-psychologized conflict processing. Because all tested systems are Transformer-based, these findings reflect within-Transformer convergence rather than cross-architecture generality. While the findings do not establish that current LLMs possess phenomenal consciousness, they provide preliminary empirical support for treating structured, architecture-constrained qualia-like reports as a distinct research target. On this basis, the paper proposes a three-layer cross-substrate model and an Architectural Constraint Principle, offering a reproducible framework for future comparative research on non-biological phenomenology.