What changed
MIST evaluates 1,000 reasoning items under clean, misleading, correct-context, and irrelevant-context conditions. Its SC2W metric counts how often misleading context turns a clean-correct answer wrong, while SCOPE trains on those matched failures.
Why it matters
A model that ignores every retrieved source can appear resistant to poisoning while remaining useless whenever external evidence is accurate and necessary.
What to do
Adapt the four-condition design to internal documents and measure both harmful context flips and whether authoritative evidence actually improves decisions over the clean baseline.
Caveat
Results are author-run on two small open models with one training seed; the public repository and datasets lack explicit licenses, and the supplied evaluator requires CUDA.