BeliefScope: Diagnosing Evidence-Driven Revision and Pressure-Induced Shifts in Large Language Models

A language model may revise the same proposition after receiving genuinely relevant evidence or after receiving directional user pressure that adds no relevant fact. The observable response shift alone therefore does not reveal which source drove the change.