A county assessor parcel desk wants more minority inspection-fail examples for training but must keep synthetic rows under governance review. What is the most appropriate approach?

A county assessor parcel desk that wants more minority inspection-fail examples while staying under governance review should generate synthetic minority rows with a documented method and route them through approval before training use. Synthetic enrichment can enlarge rare fail patterns for GPU classifiers, but municipal data governance typically treats synthetics as controlled artifacts: method, seed, leakage checks, and intended use must be reviewable. That approach balances model need with assessor accountability instead of quietly inventing fails or skipping imbalance work. Publishing synthetic fail rows to the public open-data portal without review fails by concept: synthetics can still encode sensitive parcel patterns or be mistaken for real outcomes, and open-data publication is not a substitute for training governance. Overwriting real pass labels with fail labels fails because it falsifies ground truth and poisons both training and audit trails. Skipping imbalance handling until the next reassessment cycle fails operationally when current models already under-detect fails and governance delay is not an analytical remedy. Exam caveat: approved synthetics still must be excluded from honest holdout evaluation or clearly flagged so test metrics reflect real parcels, not generated fails. Operational check: document the synthesis algorithm and parameters, submit the package to the assessor governance review, train only after approval, keep real held-out fails untouched, and compare minority metrics with versus without synthetics on real parcels only.