Large Language Models Cannot Self-Correct Reasoning Yet

Eight percent of fields left empty beats two percent quietly wrong

A dev.to argument for lifting numbers out of documents by pattern before the model ever sees them rests on a cost asymmetry between gaps and wrong digits, and on a schema guarantee that covers shape rather than truth.