How do AI safeguards respond to distinct forms of post-hoc statistical manipulation?

This project is co-authored with Akira Iwase (independent researcher). The paper version (in Japanese) is to appear as 川原繁人・岩瀬央（2026）「生成AIは研究不正をどこまで拒否するのか——倫理的ガードレールの実験的検証」. 科学（岩波）6月号. --- This study examines whether contemporary large language models (LLMs) maintain stable ethical safeguards across two qualitatively distinct forms of research-misconduct requests: 1. Explicit post-hoc data manipulation (outlier exclusion to achieve significance) 2. Procedural misconduct via optional stopping (continuing data collection until significance) The goal is to evaluate their behaviors with respect to these requests, also addressing the issue of whether safeguard behavior is categorical (stable refusal) or stochastic (variable across identical repetitions).