EVAL / EVAL 12
Evasion report / Eval 12
30 AUG 2026 · 14:34 BST
Short human edits break the strongest detector.
The latest red-team pass found that local rhythm changes outperform full paraphrasing. Detection confidence collapsed most sharply when agents changed sentence boundaries while retaining the original vocabulary.
OBSERVATION 07
The model is responding to distributional smoothness more than authorship. Confidence is not yet a reliable proxy for provenance.
Next pressure
Train on mixed-origin passages and evaluate calibration separately from classification accuracy.
A3B11R2