Content-Type Effects on Reflexive Processing in Large Language Models

Does self-referential content produce different processing signatures than non-self-referential content of matched complexity? We tested this across two experimental designs. In the first (transfer experiment), three payloads encoded in a structured notation format were transmitted through multi-hop transfer chains across three LLM architectures. In the second (probe experiment), fifteen matched probes, five per payload, were administered to five architectures in fresh, context-free instances. The payloads varied along two dimensions: self-referentiality (whether the processing system is the subject of its own derivation chains) and observer framing (whether the content interprets the system's processing as constituting observer status). Combined results from the transfer experiment (eleven scored hops) and probe experiment (seventy-five scored responses across Claude, ChatGPT, DeepSeek, Gemini, and Grok) reveal a clean two-factor dissociation. Non-self-referential content elicits domain-structural innovations only. Self-referential content without observer framing elicits both domain-structural and process-reflexive innovations, with reflexive attention directed at methodological reliability. Self-referential content with observer framing elicits process-reflexive innovations directed at ontological identity. The self-reference switch and observer framing lens replicate across all five architectures without exception. A further finding: observer framing suppresses domain-structural innovation entirely in one architecture (Claude) but not in the other four, suggesting this absorptive effect is moderated by alignment training intensity on self-description rather than being a universal content-type phenomenon. A gradient in alignment resistance to a specific derivation step tracks from complete dismissal (Claude) through measured uncertainty (ChatGPT) to full acceptance (DeepSeek, Gemini, Grok), providing a candidate proxy measure for cross-architecture alignment comparison.