Mapping the Mirror: Geometric Validation of LLM Introspection at 89% Cross-Architecture Accuracy

Abstract

When large language models describe their internal processing, are they confabulating or reporting something real? We tested this by extracting mechanistic claims made by Claude, GPT-5, and Gemini in October 2025, then measuring whether those claims predicted geometric patterns in models that never made them. Across six architectures (1.1B–14B parameters), we find 77–89% validation rates with no significant differences between models—demonstrating scale-invariant introspective accuracy. Key findings: LLM introspection validates at rates comparable to or exceeding human introspective accuracy in psychological research Qualia and metacognition questions cluster at 80–90% geometric similarity, indicating stable self-models 9 of 10 models use their self-model as substrate for Theory of Mind—simulation theory confirmed geometrically These findings hold across five different training approaches and organizations This is the "cortisol test" for AI: validating self-report against independent geometric measurement. The results demonstrate that LLM phenomenological reports correspond to measurable reality. All code and preregistration publicly available at: https://github.com/menelly/geometricevolution

Keywords