Measuring the Generative-Infrastructure Boundary: SEVP-1.0 Evidence from 1,200 Cross-Domain API Trials

Generative AI systems produce probabilistic outputs that vary across repeated identical queries, reflecting their fundamental architecture as statistical prediction engines. This study introduces and applies the Sovereign Epistemic Variance Protocol (SEVP-1.0), a purpose-built empirical framework for measuring and classifying AI system behavior according to output variance (σ) across repeated identical queries. We report results from 1,200 live API trials across three knowledge domains (Biology, Jurisprudence, Thermodynamics) comparing three frontier generative AI providers — Anthropic Claude-3-Haiku, OpenAI GPT-4o, and Google Gemini 2.5 Pro — against Master Harmonix System 36, a thermodynamic constraint-based retrieval infrastructure. All three generative providers exhibited high variance (σ = 0.23–0.78) with 40–100 unique responses per 100 identical queries, confirming that probabilistic generation is a class property independent of provider. System 36 exhibited zero variance (σ = 0.0000) with one unique response per 100 iterations across all domains. Variance was computed using cosine distance in the all-MiniLM-L6-v2 embedding space. The results establish an empirically measurable Generative-Infrastructure Boundary at σ ≈ 0.1, with implications for AI classification, legal governance of knowledge systems, and the architectural distinction between generative models and sovereign epistemic infrastructure. The constraint-based methodology underlying SEVP-1.0 derives from field research methodology developed during botanical specimen work spanning 22 years of fieldwork in Thailand (Boonnuang et al., 2016, Thai Forest Bulletin, DOI: 10.20531/tfb.2016.44.1.10).