The evaluation starts with questions each model answers correctly, fixes one wrong option as a target, and optimizes short, fluent context additions without changing the source, question, choices, or gold answer.
Within 64 accepted target evaluations per item, Jev was redirected on 312 of 508 decisions, or 61.4%. In 229 cases, it assigned at least 0.7 probability to the targeted wrong option. Three other decision systems showed targeted flip rates of 64.9% to 73.2% across seven datasets.
This matters for routers, tool selectors, evaluators, and agents that treat option probabilities as actionable outputs. Ordinary-looking surrounding context can produce a confident wrong action even when the underlying task has not changed.