Geometric Diagnostics for Continuous Style Control in Large Language Models

Explicit ratio prompts can request intermediate writing styles, but endpoint classification does not reveal proportional responses at internal ratios. We present a geometric protocol using model-specific endpoints and shared references in external sentence-embedding spaces. It excludes mechanically perfect endpoints from primary errors, retains unclipped projections, and separates axial calibration error from normalized off-axis drift. StyleBlend-Bench contains eight style axes and 320 scenarios. Five Qwen3.5 and Qwen3.6 checkpoints generated 33,600 texts over seven ratios and three seeds; three encoders produced 100,800 geometric records. An external GPT generator adds 6719 usable texts. A full five-model prompt-by-temperature experiment contains 95,994 usable condition records. Shared-reference BGE-M3 calibration error ranges from 0.176 to 0.212; other encoder/reference choices change model ordering. Two automatic evaluators correlate strongly in style intensity (r = 0.910), but content-pass agreement is 76.2%. The trajectories were rarely one-dimensional: first-component variance ranged from 0.349 to 0.408 and ordered path-length ratios from 4.64 to 5.20. The protocol diagnoses model response under specified anchors and encoders; it does not establish a human perceptual scale or absolute cross-model style coordinate.