EviCast: Forecasting Retrieval Gains Through Shared Response Calibration

Retrieval-augmented generation extends the knowledge available to large language models with external evidence. However, relevant evidence does not ensure that a generator can use it effectively. Indiscriminate retrieval can degrade task performance and increase retrieval and long-context inference costs. Existing methods neither determine before large-scale execution whether a generator, task distribution, and retriever configuration is worth deploying nor diagnose multiple candidate retrievers under a shared budget. We introduce EviCast, a framework that forecasts retrieval value through interventional evidence-state decomposition. EviCast measures a generator's response to controlled evidence states once, then combines this response profile with the offline evidence-state distribution of each candidate retriever. The resulting decomposition forecasts end-to-end net utility and its 95% lower confidence bound without executing every complete retrieval-augmented generation configuration. In frozen evaluations, EviCast correctly decides all 14 HotpotQA retrieval configurations, achieves a utility forecast mean absolute error of 2.75%, and reduces diagnostic generation calls from 480 to 96, an 80% reduction. Its topology-aware instance captures nonlinear complementarity between terminal and bridge evidence, correctly deciding all 16 configurations formed by two datasets, two generators, and four retrievers with an overall forecast error of 2.81%. These results show that retrieval deployment depends on both evidence availability and the target generator's ability to convert evidence structure into stable net utility. EviCast replaces default retrieval with a measurable, predictable, and reusable decision for resource-adaptive retrieval-augmented generation systems. Code is available at https://anonymous.4open.science/r/EviCast-ECC2.