Score retrieval sufficiency structurally before generation, not from the model's own confidence

arXiv 2609.11023·low signalXBlueskyLinkedInCopy link

RCL inserts a layer between retrieval and generation that combines a call-graph-derived structural coverage score with a novelty score estimating how far a query sits outside the model's pretraining, then triggers targeted follow-up retrieval or flags for human review when confidence falls below a calibrated threshold. The argument is that on private codebases even oracle retrieval does not stop errors, it just moves them downstream into API usage, and that model-internal confidence is the wrong signal for a private-code query. The evaluation injects synthetic internal APIs into open-source Java repos to simulate the enterprise condition.