Cosine Distance
—
Nodes
0
State
INIT
Thought Stream
⚠ Geometric Shift Detected
◈ Core Insight
By embedding each reasoning step as a vector and computing cosine distance between consecutive thoughts, geometric deviation becomes a leading indicator of hallucination — detectable before the final answer is generated.
◈ Paradox of Scale
Larger models (14B) make fewer mistakes but when they do correct, the angle errors are dramatically more severe. Small models (1.7B) stay diagnostic through frequent small corrections. Scale changes the character of reasoning geometry, not just its quality.
◈ Corrector System
Cosine distance and average angle error are monitored live during generation. When either metric exceeds a threshold, an anomaly is flagged and the task is routed to a fallback model before the output is committed.