How ClaimsText builds an evidence file

The standard it works to

A supplement may carry structure/function claims, about how a nutrient or ingredient affects the body’s structure or function, under section 403(r)(6) of the Federal Food, Drug, and Cosmetic Act and 21 CFR 101.93, with the required disclaimer. It may not claim to diagnose, treat, cure, mitigate or prevent a disease; 21 CFR 101.93(g)(2) sets out ten criteria for telling the two apart.

The claim must be truthful, not misleading, and backed by competent and reliable scientific evidence: FDA’s guidance on substantiating structure/function claims (2009), and the FTC’s Health Products Compliance Guidance (2022), which asks for the evidence to be weighed as a whole, including the results that went the other way.

1. Search

ClaimsText searches PubMed, and only PubMed: not Embase, Cochrane or other databases. Each ingredient has curated search terms in tiers: brand names, the exact form, the form's class, the parent ingredient, and bioavailability. A term must appear in a study's title or as a major MeSH topic, so the search finds studies about the ingredient rather than ones that mention it in passing.

The tiered searches are limited to clinical trials and to studies in people. Two further searches look for mechanism, which may be animal or laboratory work, and for safety and validated measures, which also admits meta-analyses and systematic reviews. There is no language or date limit. Results are ranked by how often they are cited and how recent they are, and the top records (up to 60 per report, 70 for a branded ingredient) are added to the studies already on file for that ingredient.

• Every query is shown word for word, with how many records PubMed matched and how many were retrieved.
• A query that matched more than it retrieved is marked capped, and the report says it is a documented sample of the literature, not an exhaustive review.
• If an ingredient has no curated terms yet, the report says it fell back to a search on the name alone.

2. Reading each study

A language model reads each study's title and abstract. It does not read the full text; where PubMed Central holds the full text, the report links to it. For each study it records the type and design, the number of participants, their sex, age range and health status, the product and dose given, the route, blinding, the control, and the measures used. For each outcome it records what was measured, the p-value, the effect size (copied as written, or left blank), whether significance was within one group or between groups, whether it was a primary or secondary endpoint, and whether the result went for the ingredient, against it, or nowhere.

Whether a finding counts as significant is decided by code, not by the model: a reported p-value below 0.05 is significant, and the model's view is used only where no p-value is given.

3. Is the study about this ingredient?

Every link between a study and an ingredient gets a verdict, from a language model reading the abstract: about this ingredient; about another form of it (offered only where the ingredient has a parent); one part of a multi-ingredient product; or off topic, which includes “cannot tell”. Off-topic studies, and studies with no verdict, are not cited.

A study of a multi-ingredient product can stand beside a claim but cannot carry one on its own, unless a person approves that specific link. Human studies count only where the ingredient was taken by mouth; animal and laboratory studies are exempt from that rule.

Branded ingredients have their own rule. In a report on a generic ingredient, studies of a branded version can stand beside a claim but never carry it, so a claim resting only on branded studies is removed, and the claim wording never names a brand. In a report on a branded ingredient, the same applies to studies of other brands.

4. The findings that went the other way

Every outcome is recorded, significant or not, for the ingredient or against it. The FTC treats discarding unfavorable results as deceptive; ClaimsText keeps them. The report counts the citable studies as favorable, mixed, unfavorable, null, without findings, or not read successfully, and when there is contrary evidence the executive summary includes a paragraph on how consistent the evidence is.

Only significant, favorable findings can support a claim. The memo says that the null and unfavorable studies were kept in the record and were not used to support any claim.

5. Dose fit

Code, not a model, compares the daily dose your product delivers (which you enter) with the lowest daily dose each study gave, and labels each comparison:

• At or above: your dose is at least 90% of the studied dose.
• Materially below, with the shortfall shown.
• Not comparable: per-kilogram dosing, different units, a regimen, or wording that is ambiguous. Ambiguity is never resolved by a best guess.
• Unknown: no dose was given for the product or the study.
• Each claim shows its best fit for the reviewer. Dose fit marks a claim; it does not remove it.

6. The confidence score

Each claim gets a score from 0 to 10, calculated by code from eight named components, each taking the best value among the claim's supporting studies:

• Study design: randomized controlled trial 3, observational 1.
• Blinding: double 2, single 1.
• Control: placebo 2, active 1.
• Sample size: 100 or more participants +1, fewer than 30 −1.
• Significance: between groups 2, within one group 1.
• Endpoint: a primary endpoint +1, secondary endpoints only −1.
• A validated measure +1.
• Consistency: three or more supporting studies +1.

Points come off when only an unvalidated measure was favorable, or when the closest evidence is for the parent ingredient or only about absorption, and are added for support from the brand’s own studies. A claim with no randomized controlled trial behind it is capped at 4 and marked preliminary. The reasoning for each component is written out, so a reviewer can disagree with a line rather than with a number.

7. Drafting the claim wording

A language model drafts structure/function claims only, from the significant, favorable findings of studies that may carry a claim. It is instructed never to draft a disease claim, and a word list then removes any draft that names a disease. For markers such as cholesterol or blood sugar, a claim may only speak of levels already within the normal range; it may never say a supplement lowers, raises or restores them.

Each draft then has to pass an anchor check: a language model confirms that at least one of the studies carrying the claim shows, in its own findings, the benefit the claim states. A claim the model judges only partly supported, or not supported, is removed, and the reason is recorded.

8. The disease-claim screen

Every claim is screened against the ten criteria FDA uses in 21 CFR 101.93(g)(2) to decide whether a statement is a disease claim. Code checks criteria 1 to 9 for known patterns. A language model then reviews criteria 2 to 10 for the claim, and it can only make a verdict stricter, never more lenient.

Each criterion gets its own verdict (pass, flag, fail, or not assessed) with the reason and the wording that triggered it, and the record says whether code, the model or a person decided it. A claim with any fail is marked as a disease-claim risk, and a claim with any flag is marked for review. A failing claim is not quietly dropped: it stays visible with its verdicts, so the reviewer can see why. A screen with any criterion not assessed is marked incomplete. A reviewer can record agreement or disagreement with each verdict, and the machine's verdicts are kept beside theirs.

Claims can also be checked together. A reviewer can run a net-impression check on the selected claims as a set, because claims that pass one by one can still add up to a disease claim (joints, inflammation and mobility together can suggest arthritis). If the set changes afterwards, the memo says the check was not assessed for that set.

9. Sign-off and the lock

Before signing, the reviewer records the brand, the product, the responsible person and the compliance officer. Each report carries its own, so a consultant can run reports for several client brands with their own name as compliance officer on each.

Locking a report is confirmed by typing your login email. The database records who locked it and when, from the signed-in session rather than from a form, and from then on refuses any change to the report, its claims or its studies. There is no unlock button. The memo prints the lock record and signature lines for the responsible person and the compliance officer, and where something was not assessed it says so rather than staying silent.

LabelScan

A language model reads the label: the page images, and the text layer of a PDF where there is one. Code then checks what it read against the US labeling rules in 21 CFR 101: the statement of identity and net quantity on the front panel, the Supplement Facts panel’s format, serving and daily values, ingredient and botanical names, required warnings, the disclaimer, and the business name and address. Type size and net quantity are measured from a PDF’s geometry, or from a photo and the size you give. Each health statement on the label goes through the same ten-criterion disease-claim screen as a report.

Fix and Check

An item tagged Fix was seen on the label and appears to fall short of a requirement. An item tagged Check could not be confirmed, or found, in what you uploaded: look at the printed label. Judgements a picture cannot settle, such as spelling and botanical names, are always Check. What a file or photo cannot show at all is listed separately as not assessed.

Your label

The uploaded image is deleted 30 days after a free scan; the text read from it and the findings stay with the report, and the emailed report link works for 30 days. A scan made inside a subscriber’s account keeps its label until the subscriber deletes the scan.

Where a language model judges, and where code decides

The models are called through OpenRouter, a gateway to several providers, and each judgement is stored with its reason so a person can check it. How your data is handled is set out in the privacy policy.

The limits, plainly

• It searches PubMed only, and reads abstracts, not full texts.
• A capped search is a sample of the literature, and the report says so.
• A language model can be wrong. Every judgement it makes is shown with its reason, so a person can check it.
• It covers the US rules (FDA and FTC) only.
• It is not legal advice and approves nothing. The person named on the memo chooses the claims and signs for them, and a regulatory professional should confirm anything you act on.

See it on a real product

Read the sample report, or scan your own label free.