Two made-up ads, three ways to review them. A keyword filter; retrieval over past violation cases plus an LLM judgment; and an independent verifier that re-checks every flag against its evidence. Click any line of an ad to see what each stage did with it.
Every phrase keeps its position, so a flag can be drawn back onto the original image and traced to its source — and footnotes are attached to the line they qualify.
Each line is matched against a store of past violation cases, prohibited and recommended expressions for its product category. The closest cases go to an LLM, which decides with the category's rules in front of it and writes the reason.
A separate verifier tests every flag: does the cited basis exist, does it fit the category, is the score consistent with similar lines? Flags that fail are re-judged or cleared, before anyone sees them.