I built a label checker that looks for one forbidden word pattern: disease claims

HealthPrecheck does a specific, uncomfortable job. You paste supplement or wellness marketing wording before it ships, and it returns a scored pre-check: does this read as a disease "treats / cures / prevents" claim, is a required disclaimer missing, does it imply FDA approval it does not have, and is there a substantiation cue. I built it because the gap between "supports heart health" and "treats heart disease" is where most wellness copy accidentally lands.

The line that matters most

The FDA draws a bright line between structure/function claims ("supports immunity") and disease claims ("cures colds"). The latter pushes a product into drug territory and triggers a different, heavier regime. The detection is a phrase-tier system: a blocklist of explicit disease verbs (treat, cure, heal, prevent, diagnose), a watchlist of symptom words that become disease claims in context ("lowers blood sugar" near a diabetic audience), and a check for the disclaimer that structure/function claims are supposed to carry.

This is rule work on purpose. A model will soften "this prevents infections" into "this may help your body" and call it clean. The blocklist does not soften anything. A model only restates the findings for the human. The matching is the part that has to be trustworthy, and enumeration is the part models are worst at.

An example of the context problem: "supports healthy blood sugar" is a structure/function claim and usually fine; "stops blood sugar spikes" edges toward a disease claim for a diabetic audience. The watchlist catches the second by pairing the symptom word with the audience hint, which is why I keep a human in the loop on the middle tier rather than auto-failing it. The tool reports "review this line", and a person makes the call.

What the report checks

Four axes. Disease-claim boundary: did any flagged verb or symptom phrase appear, and in what context? Missing-disclaimer check: if a structure/function claim is present, is the standard "these statements have not been evaluated" language present? Implied FDA approval: does the copy suggest endorsement the product does not have? Substantiation cues: does it invite the reader to expect proof, and is there any?

None of those is an approval. A clean report means "nothing obviously crossed the line", not "this label is cleared". The HealthPrecheck page states that plainly, because the tool's job is to stop the careless version from shipping, not to certify the careful version.

The limits I will not paper over

It cannot judge whether your underlying science is real. If you claim "clinically studied", it will not open the study. It does not know your actual FDA status, your NDIs, or your foreign registrations. It reads the words you paste and tells you where they sit relative to a well-known boundary.

It is also US-oriented — FDA and FTC framing — and does not model other countries' health-claim regimes, which can be stricter or simply different. And it cannot catch a claim that is implied by an image or a before/after photo; it is text-only.

So the output is decision-support. The report is what you bring to a reviewer or counsel so the discussion is "here are the two phrases to rewrite" instead of "is this allowed?". If a label is going into production, the flagged items deserve a real substantiation and regulatory pass. The HealthPrecheck scan just keeps the obviously over-the-line draft off the printer.

FAQ

What is the main thing it looks for?
Disease claims — wording that "treats", "cures", or "prevents" a condition, which moves a product toward drug regulation.

Does it approve my label?
No. It is a heuristic pre-check for decision-support. A clean result is not regulatory approval.

Can it verify my clinical claims?
No. It reads wording and flags risky patterns; it cannot open or assess your underlying studies.

Is it only for US rules?
It follows FDA and FTC framing. It does not model other countries' health-claim regimes.