evidence-use

A paper introducing 'Context Is Not Control', an evaluation benchmark for assessing source-boundary failures in LLMs' use of controlled text-mediated evidence. Includes replication packages for open-weight and frontier API models.