Pre-registered evaluation of sycophancy in frontier LLMs: matched-pair prompts, multi-turn pressure tests, blind human gold labels, an LLM judge validated at κ=0.89, and a system-prompt mitigation verified against true-premise controls.