Minimal fresh Reference-LoRA holdout design

Observed 2026-09-07 in e3-solutions/ender-inference, COR-3791 (branch vansh/cor-3791-fine-tune-a-freight-turn-detector-for-identifiers-pauses-and): a fresh qualification set should not be another deterministic subsample of a repeatedly used reviewed authority. Use a fixed post-deployment production window, select without model/route/outcome visibility, enforce one checkpoint per call and caller-disjointness from all training and prior holdouts, and review at most 80 proposals to seal 64 real examples (32 finished, 32 unfinished) across partial-number, complete-number, correction, pause/resumption, acknowledgement-led, and short-control strata. Compare base, selected adapter, and candidate on that same sealed set; do not compare raw scores across different holdouts.