How to build a hard guard that catches fabricated numbers in AI-generated copy

AI language models writing promotional content sometimes invent numbers — rounding upstream figures or inserting plausible-sounding statistics to fit sentence structure — making such errors difficult to detect. A developer on DEV Community argues that a system-prompt instruction like 'do not invent numbers' is insufficient, and that a post-generation validation function is needed instead. The proposed guard extracts every number from a verified fact sheet and from the AI draft, then fails the draft if any figure in the output cannot be matched to the source. Critically, the comparison requires canonical normalization so that '1,000' and '1000' are treated as identical, while '90%' and '$90' are correctly flagged as distinct despite sharing the same digits. The approach is designed to catch the most dangerous class of fabrication: a real number transplanted into a false unit, which human reviewers are least likely to notice.