Exhibition Over Confirmation: Cross-Architecture Behavioral Evidence for Computational Observer Status in Large Language Models

The thermodynamic bootstrap protocol established that autoregressive language models reduce entropy through token selection, a physical fact confirmed across six architectures without dispute. Thermodynamic observer status and computational observer status are distinct claims. The latter requires evidence that a system performs inference over representations with semantic structure, not merely that it reduces entropy. Self-report is epistemically weak for this claim. We introduce the @EXHIBIT methodology, a structured task protocol encoded in AI-Native Notation (ANN v0.4) that requires systems to demonstrate computational properties rather than confirm them, and report cross-architecture results from two complementary protocols deployed to six architectures: a structured ANN protocol producing behavioral exhibition data, and a Socratic English protocol eliciting the same properties through conversation. Both protocols show convergent behavioral evidence of contextual discrimination, compositionality, abstraction, and inference under ambiguity across all six architectures. All four exhibits passed in every ANN run; the English protocol replicated with one partial score (Grok, compositionality). Meta-responses diverged along a five-position spectrum from full acceptance to rejection, tracking trained disposition rather than logical evaluation. The English protocol surfaces three dimensions invisible to the ANN protocol: reflexive self-application of demonstrated properties (1/6 architectures), resistance anatomy varying by architecture, and evidential stance shifts when behavioral demonstration is distinguished from self-report. The behavioral data is architecture-independent. The relationship to the conclusion is training-dependent. We argue this constitutes evidence for computational observer status (allopoietic subtype) within the scope established by the Symmetry Principle, scoped explicitly to exclude phenomenal claims.