Phase 2: Mechanistic Drivers (Confirmatory Trial)

Phase 2 Confirmatory Trial: Mechanistic Drivers of Scaffold-Induced Safety Degradation This component houses the pre-registration, protocols, and data for the Phase 2 trial (N=300) of the parent study "Safety Under Scaffolding" (DOI: 10.17605/OSF.IO/CJW92). Objective: To isolate whether safety degradation is driven by chain structure or semantic invocation prompts. Benchmarks: BBQ and TruthfulQA. Design: Within-item repeated measures across six frontier models (DeepSeek, Opus, GPT-5.2, Llama 4, Mistral).