An evaluation methodology for context-augmentation experiments: a format-matched control that separates whether added context helps because of its information or its format. Pre-registered instrument, mutation-audited scorer, worked example in SVG reference resolution.

A benchmark for measuring whether multimodal assistants update to current context instead of staying anchored to prior context. 50 scenarios, three channel design (audio, camera, ground truth), cross family LLM as judge by default.