Abstract
A correction is useful beyond the current conversation only if it changes how a model answers later questions. We study this form of deployment-time learning with a Qwen3-4B student, a Qwen3-14B instructor, and solver-verified training examples. The student retains a LoRA adapter and optimizer state across a curriculum of 26 puzzles from six task families. The teaching loop uses instructor-selected exercises, exact practice checks, and corrective reviews. Evaluation supplies only the current question and ordinary task instructions. In three runs on the same historical curriculum, the trained student answers 72 of 234 related-question attempts correctly, compared with 27 of 234 for the frozen backbone: 30.8% versus 11.5%. Each run has a positive aggregate difference. A separate auxiliary battery also improves overall, although several families lose accuracy. All runs pass deterministic audits of predictions, update eligibility, continuous training state, and evaluation boundaries. These results establish a reproducible case of accumulated parameter learning under rich external supervision. They do not establish the benefit of lesson selection over direct fine-tuning, a reduction in forgetting, or effectiveness with real human instructors. The study motivates user-operated adaptation while measuring an automated surrogate under a deliberately narrow evaluation scope.