Multi-Head Behavioral Detection in Large Language Models via Fiber-Projected Hidden State Analysis

We present a multi-head behavioral detection system for large language models that achieves 125x-272x separation between target and non-target behavioral tokens using fiber-projected hidden state analysis with per-token labeling. Lightweight probing classifiers operating on 16-dimensional fiber projections extracted from frozen transformer hidden states detect four behavioral categories — repetition (125x), hedging (168x), verbosity (272x), and sycophancy (218x+) — at 0.003% computational overhead. The key methodological contribution is per-token behavioral labeling, which produces training signals 100-200x more effective than sequence-level labeling. All results are reproducible on a single consumer GPU (RTX 3090) using publicly available base models. Extended training schedules (25,000-35,000 steps) reveal continued logarithmic improvement with no overfitting. Geometric analysis confirms behavioral tokens occupy regions of 1.54x higher representational curvature in the projected fiber space.