Confidence estimated from graded past episodes beats 10-sample self-consistency at a tenth the cost.

XConf breaks the premise every existing confidence estimator shares, that confidence is readable off the current inference through introspection, token probabilities or resampling. It stores graded past episodes holding the task, reflection, stated confidence, outcome and a lesson, recalls similar tasks met with similar stated confidence to extract real success rates, then has the model name its recurring failure pattern and revise. Nine benchmarks, four models, three families: matched or beat ten-sample self-consistency on AUROC in 23 of 24 comparisons, about a tenth the ECE and a tenth the generation cost, with selective prediction raising delivered agent success up to 8.7 points. Anyone routing work by confidence should read this one.