Public-health staff find silent row corruption in a vaccination-training extract before a scheduled model refresh. What should they prioritize?

Finding silent row corruption in a vaccination-training extract before a scheduled model refresh means staff should protect training-set integrity and remediate corruption before updating the civic model. Retraining on corrupted rows can encode false coverage patterns and harm public-health decisions that depend on accurate vaccination analytics. Protecting training-set integrity and remediating corruption after updating the civic model fails because it reverses the required order; retraining first still encodes false coverage patterns from silently corrupted rows into the refreshed civic model, so later remediation cannot undo the bad baseline. Protecting validation-set integrity and remediating corruption before updating the civic model fails because it targets the wrong dataset; the silent corruption is in the vaccination-training extract, so cleaning the validation set leaves corrupted training rows intact to teach the model false coverage patterns. Protecting training-set integrity and documenting corruption before updating the civic model fails because documentation alone does not repair rows; logging the corruption without remediating it still lets bad vaccination data become the baseline for the refreshed civic model. Exam caveat: "silent" corruption may be partial; validate whether production inference already used tainted batches and whether downstream dashboards need correction notices. Operational check: halt the refresh, restore from a known-good extract or repair rows with source-of-truth systems, recompute checksums, document the incident, and only then rebuild and revalidate the vaccination model.