How do you control false-call and escape rates?

False calls and escapes are a trade-off controlled at four levels: imaging quality first, using suitable illumination to raise defect contrast and reduce algorithm burden at the source; multi-feature voting, cross-validating each verdict against several features to reduce single-point errors; configurable thresholds and confidence, so the line can alarm below a lower limit and reject above an upper limit; and calibration with real defective parts including boundary samples, with results written into acceptance criteria. On real projects we run a measured set with your good and defective parts to report detection and false-call rates rather than a generic promise.