Filed under training-data.

The Defect You Cannot Photograph

Industrial AI vision inspection vendors cite 97–99% detection accuracy — a figure almost always trained on defects the line never made. Synthetic data can cover rare failure modes, but only after someone characterises the production environment precisely enough to generate physically plausible images and then audits the output by hand. The months nobody budgets for that work are what separate a convincing pilot from a system that holds under real production variance.