RCL scores retrieval sufficiency structurally before generation, not from the model's confidence.

arXiv 2609.11023 inserts a layer between retrieval and generation combining a call-graph-derived structural coverage score with a novelty score estimating how far a query sits outside pretraining, then triggers targeted follow-up retrieval or flags for human review below a calibrated threshold. The argument is that on private codebases even oracle retrieval doesn't stop errors, it moves them downstream into API misuse, and model-internal confidence is the wrong signal for a private-code query. Evaluated by injecting synthetic internal APIs into open-source Java repos.