An assessor’s office feeds features into a valuation model. Which security control best protects that pipeline?

Implementing input validation and schema checks before valuation features hit the model is the security control that protects an assessor's AI valuation pipeline: unvalidated inputs enable data poisoning, injection of out-of-range features, and garbage scores that still look authoritative. The correct control works by enforcing types, ranges, required fields, and referential integrity at the boundary so only well-formed feature vectors reach the model. Logging valuation feature inputs for audit review before valuation features hit the model fails as a pipeline control because it is detective rather than preventive; it records features for later review but cannot block or reject malformed or poisoned features before inference, so authoritative-looking bad scores can still be produced. Encrypting valuation features in transit before valuation features hit the model service fails to protect the pipeline because encryption only secures data in motion; it does not validate types, ranges, required fields, or referential integrity, so malicious or malformed feature vectors can still reach the model inside encrypted channels. Enforcing role-based access on feature intake system before valuation features hit the model fails to protect the model pipeline because it governs who can submit to intake, not whether submitted features are well-formed; an authorized user or compromised account can still send malformed or poisoned features that produce a bad inference in the first place. Exam caveat: validation rules must track feature-schema version changes when the model is retrained, or checks will reject legitimate new fields or miss newly critical constraints. Operational check: define a versioned feature schema, reject nonconforming records with logged errors before inference, alert on validation-failure bursts, and include schema-validation tests in the valuation pipeline's CI promotion path.