The pristine ARRL Code Practice corpus is too clean to train against — it collapses adaptive σ floors and pulls decision boundaries inward, regressing real OTA performance. This experiment delivers a ...