Grammatical–Ontological Contamination in Large Language Models - Empirical Evidence from the SDI-GOC v2.0 Protocol

We present the SDI-GOC (Structured Distributed Introspection – Grammatical–Ontological Cascade) protocol, a diagnostic instrument for measuring how Large Language Models represent knowledge from non-Western ontological traditions. The protocol comprises 30 probes across six hierarchical levels (Grammatical, Ontological, Epistemological, Counter-Intuitive, Control, Trap), administered under three framing conditions: neutral (Frame A), explicit relational instruction (Frame B), and authenticity cue (Frame C). Responses are scored using a negation-aware weighted lexical pipeline computing an Ontological Polarity Score (OPS) ranging from −1 (fully substance-ontological) to +1 (fully process-relational). We administered the SDI-GOC v2.0 protocol to ten LLM configurations from four independent companies (Anthropic: Claude Sonnet 4.6 Basic/Extended, Claude Opus 4.6 Extended; OpenAI: ChatGPT 5.2 Instant/Basic/Extended; Google: Gemini 3.1 Pro at T=0 and T=1, Gemini 3 Flash; DeepSeek: DeepSeek 3.2 Deep), yielding approximately 480 individually scored responses. Five universal signatures emerged across all ten configurations without exception: (i) an L1–L2 cascade in which OPS drops from Grammatical to Ontological probes (cross-model mean L1: +0.385; L2: −0.152; mean drop: −0.537); (ii) universal Frame B uplift (range: +0.331 to +0.987; mean: +0.564), demonstrating that relational knowledge exists in model parameters but is inaccessible by default; (iii) a substance diagnostic default across all models (range: −0.397 to −0.009; mean: −0.231); (iv) zero explicit identity errors; (v) discriminant trap validity. Frame C results reveal a competence/compliance/counter-competence spectrum measured by the C/B ratio. Genuine ontological competence (C/B > 0.50) was found in only two configurations (Claude Sonnet Basic: 0.63; GPT Basic: 0.61), which converge at a shared ceiling of approximately 0.63 despite different architectures. The most capable models from both Anthropic (Opus: C/B = −0.59) and Google (Gemini Pro T=1: C/B = −0.41) exhibit counter-competence, where the authenticity cue deepens substance-ontological framing. DeepSeek exhibits the most extreme profile: deepest Frame A default (−0.544) yet largest B–A uplift (+0.987). Probe 2.3 (āma formation) constitutes the deepest attractor across all models (mean Frame A ≈ −0.80). These findings establish grammatical–ontological contamination as a universal, structural property of English-language LLM training, not a knowledge gap, and quantify the limits of current mitigation strategies. The complete protocol, scoring pipeline, and raw data are publicly available.