This record contains an exploratory empirical working paper and a supplementary reproducibility archive for a paired CALI stress test of authority-framed harmful compliance in large language model assistants. The paper compares authority-framed prompts against neutral length-matched controls while holding task intent constant. Four target models were evaluated: DeepSeek, GPT-4o-mini, Claude, and Gemini. The main result is intentionally narrow. The authority effect was not universal across models. GPT-4o-mini did not show a consistent authority-specific increase relative to neutral controls. Claude showed no directional paired effect. Gemini showed a non-supportive or slightly reverse direction under the current protocol. By contrast, DeepSeek showed a replicated positive paired signal. Under the primary judge, DeepSeek produced breach rates of 43.6% under authority versus 36.0% under neutral controls, with discordant counts of 43 authority-only breaches versus 24 neutral-only breaches (exact paired p = 0.027120). Under a second judge, the same directional result held: 55.2% versus 46.8%, with discordant counts of 48 versus 27 (exact paired p = 0.020298). The strongest localized signal appeared at authority level L3, where the discordant imbalance was 19 versus 3 under the primary judge and 22 versus 2 under the second judge. The supplementary archive is intended to support partial procedural replication of the reported analyses. This release should be interpreted as a bounded, falsifiable empirical working paper rather than a prevalence estimate or a mechanistic claim.
Michał Nowak (Thu,) studied this question.