PreFlight - YA

PreFlight screens health and food-supplement adverts against UK advertising rules before they are published. In the UK, a supplement ad may make a health claim only if the GB nutrition and health claims register authorises that claim for that substance, and only in wording that means the same thing. The CAP Code adds rules on medicinal claims, testimonials and evidence, and the Advertising Standards Authority (ASA) rules on ads that break them, publishing each upheld ruling under the advertiser's name. With the rules spread across a legal register, a code of practice and a long record of rulings, breaches are easy to miss, and they are usually found in public, after the ad has run. PreFlight takes the ad in whatever form it already exists, whether a video, an image, a carousel, a product page or the script, and lists every claim it makes, including those carried by a picture, such as a before-and-after shot, or by a testimonial. Each claim is checked against the register, the CAP Code and similar past ASA rulings, and every flag shows both where the claim sits in the ad (a transcript timestamp, a region of a frame or a numbered script line) and the rule or ruling it conflicts with. Where the register authorises a claim for the same substance, PreFlight quotes that wording exactly, with its register ID, as the replacement; it never writes replacement wording of its own. A fix loop then applies the changes, editing a still image when the picture carries the claim, and screens the ad again until nothing is flagged or a person has to take over. PreFlight gives no legal advice and never approves an ad, so a clean result reads "no issues found against" the dated versions of the sources it used, and the decision to run the ad stays with a person. We tested it on 79 ads it had never seen: 49 rebuilt from upheld ASA rulings, 10 clean ads with one breaching line added, and 20 clean ads that use only authorised wording. Since the ruling pages describe each ad but publish no image or video, each ruled ad was reconstructed from the ruling's own description. PreFlight flagged 54 of the 59 ads with a breach (92%) and 1 of the 20 clean ads, against 28 of the 59 for a keyword baseline, and at the level of single claims it caught 63% of breaching claims with a matching rule where the baseline caught 9%. PreFlight runs as a web workspace, as a command-line tool, and as two tools that other AI agents can call to screen an ad before they publish it.