Grammatical–Ontological Contamination in Large Language Models: Corrected Empirical Evidence from the SDI-GOC Protocol across Ten Configurations and Four Companies

We present the corrected results of the SDI-GOC (Structured Distributed Introspection – Grammatical–Ontological Cascade) protocol, a diagnostic instrument for measuring how Large Language Models represent knowledge from process-relational medical traditions. The protocol comprises 30 probes across six hierarchical levels, administered under three framing conditions to ten LLM configurations from four independent companies (Anthropic, OpenAI, Google, DeepSeek), yielding 472 individually scored responses. Two findings remain after correction. First, a substance-ontological diagnostic default appears in all ten configurations. Second, the non-instructed L1–L2 lexical cascade appears in 10 of 10 configurations, with a canonical mean drop of −0.841 OPS points (95% CI [−1.062, −0.603]). The v2.0 value was diluted by pooling instructed and non-instructed responses; the v3.0 value was further affected by substring matching in its negation filter. A structural analysis of the same 472 outputs separates terminology from clause construction. Between the neutral and relationally framed conditions, lexical polarity differs by +0.605 and structural polarity by +0.038. The interaction is +0.566 (95% CI [+0.459, +0.686]), positive in 10 of 10 configurations. Structural scores remain negative in every configuration under every framing condition. Across the 29 actual uploaded-document sessions, the two indices correlate at r = 0.668 (configuration-cluster bootstrap 95% CI [+0.544, +0.791]). The condition contrast is substantially larger lexically than structurally, but item differences prevent causal attribution to instruction alone. The structural metric is exploratory and unvalidated against blind human annotation. Three v2.0 claims remain withdrawn. The Frame B uplift is confounded by prompt lexicon and item non-equivalence; 79.6% of recovered Frame B process-marker weight under the original audit derives from terms present in the prompt. The C/B ratio and its three-zone classification are unstable under leave-one-probe-out jackknife. The only lexically unprimed test, Frame C minus Frame A on matched probes, yields +0.224 with a 95% CI of [−0.008, +0.465], positive in six of nine configurations; the interval includes zero and does not establish latent relational capacity.