Caption Audit

Audit a LoRA caption set for tokens fused with the trigger word. Point 'directory' at a folder of images + same-named .txt captions; the audit runs there and the report card renders inside the node. The primary measure is document frequency — in how many captions a term appears at least once, across words, phrases and whole comma segments. A term near 100% cannot be prompted in or out at inference: the model cannot tell it apart from the trigger. This reads .txt files only, never the images. Every flagged term has two possible causes needing opposite fixes, so the card ends with the question you have to answer by looking at the pictures. Wire 'critical' into a gate to stop a training workflow before it starts.