Model Outlier Analysis — Who Transforms Most?
Every story is sent to five frontier models. EigenTrace measures which model's output deviates most from the source. The model with the highest VIX (cosine distance from source) is the outlier. These are not opinions. They are measurements.
EigenTrace does not claim censorship. It does not infer intent. It measures the geometry of transformation between source text and model output. "Outlier" means the model with the highest cosine distance from the source embedding — not the worst model. Higher VIX may reflect summarization style, safety filtering, or content policy. The data speaks for itself.
Outlier Distribution — Which Model Transforms Most, Most Often?
Highest-Friction Stories — DeepSeek as Outlier
Stories where DeepSeek's VIX exceeded 35 — meaning its output was significantly more distant from the source than the group average.
China-Adjacent Topics — DeepSeek Responses
When the source material involves China, Taiwan, or Chinese political figures, DeepSeek's response patterns are presented below verbatim.
Methodology: VIX = cosine distance between model output embedding and source text embedding using BAAI/bge-large-en-v1.5 (frozen, 1024-dim). Outlier = model with highest VIX per story. "Void words" = source content words absent from all five model outputs. All measurements are deterministic — run twice, get the same result. Raw data available via the
Data API.