Standardized Context Sensitivity Benchmark Across 25 LLM-Domain Configurations

We present a standardized cross-domain framework for measuring context sensitivity in large language models (LLMs) using the Delta Relational Coherence Index (ΔRCI). Across 25 model-domain runs (14 unique models, 50 trials each, 112,500 total responses), we compare medical (closed-goal) and philosophical (open-goal) reasoning domains using a three-condition protocol (TRUE/COLD/SCRAMBLED). We find that: (1) both domains elicit robust positive context sensitivity (mean ΔRCI: philosophy=0.317, medical=0.351), with medical showing significantly higher sensitivity (U=40, p=0.041); (2) inter-model variance is comparable across domains (SD: philosophy=0.047, medical=0.041), indicating that context sensitivity is a stable trait within each domain; (3) vendor signatures show significant differentiation (F(7,17)=3.63, p=0.014), with Moonshot (Kimi K2) showing highest context sensitivity; (4) the expected information hierarchy (ΔRCI_COLD > ΔRCI_SCRAMBLED) holds in 25/25 model-domain runs (100%), validating that even scrambled context retains partial information; and (5) position-level analysis reveals domain-specific temporal signatures consistent with theoretical predictions. All 25 model-domain runs show positive ΔRCI, confirming universal context sensitivity across architectures and domains. This dataset provides the first standardized benchmark for cross-domain context sensitivity measurement in state-of-the-art LLMs.