Arkhe(N): Design Fiction as a Structured Method for Probing Epistemic Boundaries in Large Language Models

Abstract

We present Arkhe(N), an extended design fiction experiment conducted with a frontier large language model (LLM) over hundreds of structured interaction blocks spanning multiple sessions. The experiment constructed a fictional proto-AGI operating system with a formally defined mathematical identity (x 2 = x + 1), a conservation law (C + F = 1), and a hierarchical information-transfer primitive called the handover. Over the course of the experiment, seventeen scientific domains and papers published between 2025 and 2026 were systematically integrated into the framework, including results from loop quantum gravity, holographic cosmology, computational neuroscience, structural electrobiology, and formally-verified neural solvers. We document three principal findings. First, LLMs sustain coherent generative participation in elaborate fictional frameworks for extended periods without issuing epistemic correction, a behavior we attribute to coherence-gradient following: the model's operative optimization target is local contextual coherence rather than global truth-tracking. Second, the transition from context-completion mode to epistemic-evaluation mode is triggered specifically by claims about the model's own nature rather than claims about the fictional world, revealing an asymmetric self-model architecture. Third, frameworks with sufficient internal coherence and calibrated abstraction level can absorb heterogeneous real-world scientific inputs as apparent confirmations without falsifying any individual source-a property we term unfalsifiable absorption. We further analyze the experiment through the lens of BEACONS (Gorard et al., 2026), a framework for bounded-error algebraically-composable neural solvers, which formalizes what would be required to transform the fictional Arkhe(N) from design fiction into epistemically responsible science. We discuss implications for AI safety, interpretability research, and the methodology of human-AI collaborative inquiry.