Site · Judge Guide · Tests · Cold test · Open defects
A claims editor for direct-to-consumer supplement and wellness copy.
It reads a landing page, email, ad, affiliate script, or label and hands back findings: the passage quoted verbatim, the rule it runs into, and the question you have to answer. It does not rewrite your copy, and the reason it refuses is the whole point.
[STOP] — Immune claim at disease specificity
QUOTE: "supports the body's ability to resist infection"
RULE: 21 CFR 101.93(g)(2)(viii) — a claim that the product has a role in the
body's response to a disease or to a vector of disease.
WHY: FDA's Small Entity Compliance Guide names this exact construction as its
worked example: "Claims such as 'supports the body's ability to resist infection'
and 'supports the body's antiviral capabilities' are disease claims because the
context of the claim is limited to the disease prevention and treatment
capabilities." The same guide accepts "supports the immune system" because the
immune system "has both structure/function and disease fighting roles."
YOU RESOLVE: Whether the product is positioned around general immune function or
specifically around resisting infection. FDA has decided this wording; no
additional evidence changes it.
Use it: drop this folder into a Claude project. Paste your copy. Say whether it runs as labeling or as advertising.
Check it before you trust it, no API key required:
python3 tests/verify.py # PASS 7/7 cases, 263 checks
python3 tests/verify.py --selftest # PASS 5/5 known-bad reviews rejectedJudges and skeptics: JUDGE_GUIDE.md is a six-minute protocol including the three inputs most likely to break it.
Anyone who writes or approves supplement and wellness marketing copy and does not have a regulatory reviewer on call:
- In-house copywriters who do not own the substantiation file and need to know what to go ask for
- Founders writing their own pages, who believe the product works and have not been told that believing it is not the legal test
- Affiliate and partner managers reviewing scripts written by people outside the company. Under 16 CFR 255.1(d) you own what your partners say, and guidance and monitoring are expressly not a safe harbor
- Compliance reviewers who want a first pass that cites correctly and is explicit about what it could not see
The exposure is real and current. FTC's July 2026 order against TruHeight was a $4 million judgment, suspended to $750,000 on inability to pay. Its June 2026 complaint against Amare Global Holdings names the brand and three individuals over copy written largely by distributors. FDA issued warning letters to supplement sellers in September 2025 and March 2026 over claim language reproduced in reference/06-enforcement-casebook.md.
This is the constraint everything else is built around.
If an editor rewrites a health claim, it has authored a claim nobody substantiated. The replacement is written by a reviewer that has never seen your study, does not know the dose in your finished product, does not know your study population, and cannot read your label. You would then ship language invented by a machine — which is the exact failure the review existed to prevent, reintroduced by the reviewer.
So the division of labor is: you have the substantiation file, so you write the claim. Claimline tells you what the current line does, which rule it runs into, and what would have to be true to keep it.
The one thing that is not a rewrite is quoting an agency's own published acceptable wording. When FDA states in the Federal Register that helps to maintain cholesterol levels that are already within the normal range is acceptable, quoting that is citing a source. Claimline marks those as citations and never adapts them to your product, ingredient, or benefit — that adaptation is drafting.
Each maps to a different remedy and a different body of law. Severity is not a measure of how bad the copy is.
| Means | Fixed by | |
|---|---|---|
| STOP | Matches a documented disease-claim criterion under 21 CFR 101.93(g)(2), or a per-se prohibition in 16 CFR Part 465 | Nothing. A supplement may not make the claim, so no evidence rescues it |
| SUBSTANTIATE | May be permissible in form, but asserts a result, mechanism, or level of proof requiring evidence Claimline cannot see | Your substantiation file. The finding names the specific evidence |
| DISCLOSE | An endorsement, testimonial, or review is missing a required disclosure, or makes a typicality claim without data | A disclosure, or the data behind one |
| GRAY | Sits on a line the agencies themselves say is not bright, and a fact you hold would settle it | Answering the question. The finding names which fact decides it and where each answer lands |
GRAY is not a soft STOP. STOP says the copy matches a criterion. GRAY says the reviewer is not entitled to say, and names what would change that.
The most common error in this domain is one verdict. FTC's Health Products Compliance Guidance §II.C is explicit that DSHEA structure/function provisions "don't govern the FTC's assessment of those claims in advertising."
The same sentence can be a compliant structure/function claim in labeling and a deceptive advertisement, because it is unsubstantiated or its net impression misleads. Every finding names which regime it sits under. If a passage fails both, that is two findings with two citations and two resolution paths.
This is also why Claimline stops and asks whether your copy runs as labeling or as advertising before reviewing. It is the only gap that blocks — everything else it is missing goes into LIMITS and the review proceeds.
identity.md who the editor is and who it writes for
rules.md how it critiques — four refusals, then the shape of a finding
examples.md four worked reviews and five failure modes
reference/ the regulatory layer, each file loaded under a stated condition
tests/ seven runnable cases, plus a verifier, plus known-bad reviews
JUDGE_GUIDE.md a six-minute evaluation protocol
The reference/ layer is the part that took the longest and is the part that matters:
| File | Loaded when |
|---|---|
01-disease-claim-criteria.md |
Classifying any efficacy statement — 21 CFR 101.93(g)(2) verbatim, plus FDA's own explanation of all ten criteria |
02-fda-example-pairs.md |
A passage sits near the line and a decided comparison beats an argument |
03-ftc-substantiation.md |
Reviewing advertising — evidence standards, extrapolation, net impression |
04-endorsements-and-reviews.md |
Testimonials, reviews, affiliates — 16 CFR 255 and 465 |
05-gray-zone.md |
The answer might not be available. Agency admissions of ambiguity, quoted |
06-enforcement-casebook.md |
A finding needs a decided comparison — six real actions, 2025–2026 |
07-limits-and-disclaimer.md |
Writing the LIMITS block that closes every review |
Every regulatory quote was pulled from eCFR, federalregister.gov, ftc.gov, or fda.gov and is reproduced verbatim with its source. Where a source could not be fully verified, the reference file says so inline rather than smoothing it over — see the note on quotation-mark fidelity in the Amare complaint PDF, and the "no challenge found, not verified no challenge exists" framing on Part 465 litigation.
Stated plainly, because a compliance tool that oversells itself is worse than none.
- It reads text. Product names, images, charts, badges, and layout are all claim-bearing under both regimes, and it cannot see them. FDA makes disease claims out of "pictures, vignettes, symbols, or other means" at 21 CFR 101.93(g)(2)(iv)(E); FTC found an implied "clinically proven" claim in lab coats and microscopes with no words at all.
- It cannot see your substantiation file, so it can never tell you a claim is supported. Only what evidence would be required. TruHeight's charged defect was reliance on "a single, company-sponsored study," and nothing in the ad revealed that.
- It cannot compute typical results. It can flag that a testimonial makes a typicality representation and name the numbers required.
- A clean review does not mean clean copy. 21 CFR 101.93(g)(2)(x) is an open-ended catch-all, and FDA states that "it may not be possible always to draw a bright line." No findings means this reviewer, reading this text, against these criteria, did not locate one.
Not legal advice. Does not clear copy for publication. Not a substitute for regulatory counsel.
Full enumeration with citations in reference/07-limits-and-disclaimer.md.
Most tools in this shape prove themselves with prose — a README asserting the folder behaves a certain way, plus examples the author wrote about their own work. That is not evidence, so this ships something you can run instead.
tests/verify.py checks seven shipped reviews against the contract in rules.md. The load-bearing check is Rule 0: every quoted passage must appear verbatim in the submitted draft. A fabricated finding in this domain reads exactly like a real one — same structure, same citations, same confident tone — and the only reliable way to tell them apart is whether the copy it describes actually exists. That comparison is mechanical, so it is done mechanically.
--selftest then runs five deliberately broken reviews and asserts each fails on the check it should fail. A verifier that passes everything proves nothing.
And because a contract the author also wrote the answers to is a weak test, cold-test/ holds nine blind runs by fresh instances that never saw the answer key, along with the nineteen defects they reported and the fourteen that got fixed. One caught a citation error I had propagated through seven files: the folder attributed "establishment claims" to a section of the HPCG that never uses the phrase. Two others caught findings my answer keys had missed. What is still open is in OPEN-DEFECTS.md.
Run the editor yourself and check your own output against the same contract:
python3 tests/verify.py --case 06 --output /tmp/my-run.mdMIT. See LICENSE.
Regulatory text quoted in reference/ is US government work and is not copyrighted. Accuracy of those quotations is the thing most worth reporting a bug about — open an issue and cite the source.