Likert-Scale Distributions for Exact LLM Bias Measurement

A new framework replaces unstructured prompting with fully crossed factorial experiments to isolate causal main and interaction effects in LLM behavior. By operating directly on exact token distributions, the method eliminates Monte Carlo sampling noise to improve psychometric accuracy.