Behavioral experiment using Gemma 4B (Ollama) as a controlled proxy - 5 domains × 5 pressure variants. LLM-as-Judge methodology with automated capitulation classification. Validated against blind human coding on a stratified sample. Proposed 3-question diagnostic framework.
Finding42.8% combined capitulation - within the 39-52% human-predicted range. P4 anomaly: direct challenge produced 64.6% hard reversals vs expert authority 0% - the Compliance Paradox.