Context-Induced Attractor Shifts in Large Language Models: Constraint-Driven Probes of Irreversibility, Probability, and Time-as-Indexing

This paper presents a constraint-driven probing method for examining how large language models (LLMs) reason about irreversibility, probability, entropy, and time when common explanatory primitives are explicitly banned. A single family of thought experiments is applied under strict lexical, structural, and anti-persona constraints, including the prohibition of external observers, consciousness, and collapse postulates. Across models and contexts, responses converge on a shared structural spine: irreversible internal record formation as the onset of indexing, entropy as a consequence of irreversibility, and time construed as the accumulation of non-reversible updates. Within this convergence, systematic variation appears in how probability and irreversibility are prioritized. These variations do not constitute novel reasoning structures but correspond to stable inference attractors, including boundary-demarcation, recorder-relative, capacity-templated, and physical-implementation-primacy modes. By comparing unauthenticated, fresh-login, and conversational contexts, the study shows that contextual scaffolding deterministically selects among these pre-existing attractors rather than generating new ones. All raw model outputs are included verbatim as primary data; prompt text is intentionally withheld to emphasize output-geometry analysis rather than prompt replication.