Check training exposure before expanding a fine-tuning dataset

When a small fine-tune misses examples it was trained on, audit training fit and effective exposure before concluding that more data is the next fix. At fixed optimizer steps and batch size, increasing the dataset reduces average passes over each example; a comparison that changes both corpus and exposure cannot isolate a data benefit. First check labels, assistant-token masking and serialization, then test one predeclared longer training horizon with identical data, base weights and inference settings. Keep training-fit diagnostics separate from held-out evaluation, compare paired fixes and regressions by failure family, and retain failures in the denominator.

Limits: imperfect training fit motivates an optimization hypothesis; it does not prove undertraining, and lower training loss or improvement on authored tests does not establish real-call prevalence or production improvement. Repeatedly inspected regression sets are exposed tests, not fresh validation. Inspect changed-wrong cases as well as correct-to-wrong regressions, because switching from a wrong identifier to clarification can change downstream retry behavior.

Evidence/example (2026-09-10): inspection of ender-inference training/mc_pilot/train.py and saved training receipts found fixed-step runs over different corpus sizes, with valid assistant masks and serialization. This motivated a same-data, one-variable optimization-horizon experiment and separate authored versus exposed-regression reporting. No customer records, raw call data or deployment-state claims are included. This method extends the cited advice to freeze inference settings and inspect paired errors.