Default-filled outcome labels in a deployed cognitive-screening programme: an operator-level audit and the construction of twenty-four language-model arms

Background. Routine service databases are attractive sources of training labels for clinical prediction models, but the processes that write those labels are rarely audited before the labels are used. In a deployed community cognitive-screening programme, we audited the routine cognitive-status label, built a matrix of twenty-four model arms over the same patients under a specialist reference standard, and measured what each supervision choice bought or cost. Methods. The study cohort is the 672 individuals whose cognitive status was recorded by a titled (attending-or-above) physician, that record being the reference standard; after holding out one institution entirely, a development panel of 642 individuals at 38 institutions. The routine cognitive-status label these individuals also carry was first audited at the operator level: for each data-entry account we counted diagnoses entered and the proportion recording any impairment, and tested a competing bulk-timestamp explanation. Twenty-four arms span the supervision choices such a programme faces: an incumbent 21-variable logistic regression; local language models (Qwen2.5-1.5B/3B, Qwen3-4B/8B) zero-shot, with chain-of-thought, fine-tuned on physician labels, on routine labels with and without decontamination, or on a proxy scale-band task; preference-optimised (DPO) and reinforcement-trained (GRPO) variants; a proprietary frontier model queried zero-shot; and knowledge distillation of that frontier model into the regression and into the local 4B, using 943 teacher-labelled records from the programme's unlabelled pool. All arms are scored out-of-fold under one five-fold split grouped on registry-resolved institution clusters (no cluster spans a fold); paired contrasts use a 2,000-draw cluster bootstrap. Results. 181 operator accounts (each entering at least 100 diagnoses with zero recorded impairments) account for 45,315 rows - 40.5% of the outcome column; recorded impairment falls monotonically with account volume (15.7% for 1-9 rows to 0.7% for 500-999); a bulk-timestamp explanation was tested and refuted, identifying the write-time column as a migration artefact. Under the specialist standard, no locally fine-tuned arm beat the incumbent regression (AUROC 0.926): physician-label SFT reached 0.924 (4B), DPO 0.881, and GRPO 0.789; the pre-registered two-stage proxy-then-RL recipe was worse than its single-stage contaminated baseline (-0.030, 95% CI -0.077 to -0.004). Chain-of-thought reduced discrimination at every size (-0.072, -0.080, -0.041 at 1.5B/3B/4B; -0.012, n.s., at 8B). The frontier model scored 0.932 (vs. regression +0.007, n.s.). The distilled 4B reached 0.940 - above the incumbent (+0.014, 0.004 to 0.031) and above its own teacher (+0.008, 0.001 to 0.017) - with near-teacher calibration; it reached the teacher's level by 50 teacher labels and changed little beyond 200. Conclusions. The audit and the arm matrix support one deployment recipe: audit the routine label at the operator level before training on it; do not expect fine-tuning, preference optimisation, or reinforcement learning on a few hundred specialist cases to beat a well-calibrated regression; and if a frontier model is available but undeployable, spend a bounded number of queries on it as a labelling instrument and distil. A companion paper uses these frozen predictions to quantify how evaluation design choices compare with model choice.