Language models can abandon correct answers when a "verified" source disagrees. We study how this source deference differs from agreement with a user, and whether activation interventions can reduce it. This repository contains code for behavioral evaluations, attribution patching, and steering experiments.