ClaimGate
Flag marketing claims your evidence cannot substantiate, before you publish. Checks EU AI Act Article 50 disclosure.
Documentation
Your AI wrote it.Can you prove it?
ClaimGate checks every claim in a draft against the evidence your organisation actually holds — the spec sheet, the case study, the paper. It enforces your prohibited-claims policy, and it flags the AI disclosure the law now requires. If a claim cannot be substantiated, the build fails.
Free: the starter kit — policy file, evidence folder, six-question checklist and a CI job. MIT source, no dependencies, Python 3.9+.
Claims requiring substantiation (14)
✗ [attribution] According to research at a leading university, 94% of teams…
unsupported: the figure(s) 94% do not appear in any supplied evidence
✗ [statistic] …and we are fully ISO 27001 certified.
contradicted: evidence states "…is not certified under ISO 27001"
~ [statistic] Meridian cut reconciliation from 22 hours to 9 hours
partially_supported: the case study confirms both figures
try: Meridian's finance team reduced reconciliation from 22 hours a week to 9
Policy findings (3)
✗ [prohibited] guaranteed
absolute guarantees are unenforceable and treated as misleading advertising
✗ [disclosure] AI disclosure missing
required under EU AI Act Article 50 and by platform labelling rules
BLOCKED — 14 unsupported claim(s), 3 blocking policy finding(s). (exit 1)
Use it from an AI agent
The same check runs as an MCP server, so an agent can gate a draft before it publishes instead of someone remembering to. Two entries are listed in the official MCP Registry, and neither needs an account or anything installed beyond uv: io.github.caseone115/claimgate for the full engine, and io.github.caseone115/claimgate-kit for the free starter kit, which writes the policy, the evidence folder and the CI job into your repository before it checks anything.
{
"mcpServers": {
"claimgate": {
"command": "uv",
"args": ["run", "--with", "mcp", "python", "server.py"]
}
}
}
The engine entry gives five tools: check_draft, check_draft_json, list_claims, check_policy_only and about. It fails closed the same way the command line does — with no evidence it answers “unverified”, never “fine”. Source and setup.
The gap nobody closed
The problem with AI-assisted copy is not that it reads badly. It is that a reader cannot tell a fabricated statistic from a real one — there is no signal in the prose that separates them. So it passes.
📄
2 Aug 2026
Article 50 transparency obligations apply. AI-generated content must be disclosed in defined circumstances — and by 2 December 2026 the machine-readable marking duty in Article 50(2) applies too.
Regulation (EU) 2024/1689, Art. 50 and Art. 113; European Commission transparency guidelines.
⚖️
69–88%
hallucination rate on specific legal queries, across the state-of-the-art models the study tested — and the models did not know they were wrong.
Dahl, Magesh, Suzgun & Ho, Hallucinating Law, Stanford RegLab & HAI, Jan 2024; 200,000+ queries per model. stanford.edu.
💰
$492M → $1B+
forecast spend on AI governance platforms, 2026 → 2030. The budget to keep this in order already exists.
Gartner press release, 17 Feb 2026 (Global AI Regulations Fuel Billion-Dollar Market for AI Governance Platforms). Gartner's own pages block automated readers, so this links to a page that reproduces the figure with the release named — the primary document is findable by its title and date.
Every figure on this page carries the document it came from. That is the rule this product enforces on your copy, so it applies to this page first — and a figure with no source is the exact failure mode it exists to stop.
Detectors answer the wrong question
They tell you a text was AI-written. Nobody was ever blocked from publishing over that. The question that decides whether an asset ships is whether this claim can be traced to something you can point at.
"No source, no claim" — enforced
Most style guides already say this. Nothing checks it, so under deadline pressure it erodes. ClaimGate makes it mechanical, and makes the build fail when it is broken.
Three checks, one exit code
Every finding quotes the text it is about, says what is missing, and says what to do. A gate an editor cannot argue with is a gate they turn off — so nothing here is a black box.
🔍
1. Substantiation
Every checkable claim is extracted and adjudicated against your evidence. Any number in a claim must appear in the evidence — that check is a string search, not a judgement call, so a fabricated figure cannot be talked into a pass.
📋
2. Policy
Your prohibited and restricted claims, written down once and enforced on every draft: guarantees, medical and financial claims, unsourced superlatives, named competitors, personal data.
🏷️
3. Disclosure
A customer-facing AI-assisted asset with no AI disclosure is non-compliant however good the copy is. Checked as a property of the draft, not left to whoever uploads it.
Relying on the human-review exemption? Article 50(4) exempts AI-assisted published text from labelling only where the content had human review or editorial control and a named person holds editorial responsibility — and the Commission’s July 2026 guidance says the responsible person’s identity and contact details must be publicly available. What that asks you to be able to prove, and the six questions it reduces to: the editorial exemption is an evidence trail, not a tick-box →
It fails closed. If the model is unavailable, the key is missing, or the reply is malformed, the verdict is unverified — which blocks. It never fails towards "supported". The tests assert this, because it is the one behaviour that makes the tool safe to trust.
How it compares
Adjacent tools are useful and none of them do this.
| Question | Detectors | Governance suites | ClaimGate |
|---|---|---|---|
| Was this written by AI? | Yes | — | Not asked |
| Is every claim substantiated? | No | No | Yes |
| Enforces your banned-claim list? | No | Partly | Yes |
| Checks AI disclosure obligations? | No | Partly | Yes |
| Blocks the publish, in CI? | No | No | Exit code 1 |
| Can you read the rule that fired? | No | Rarely | Every finding quotes it |
Pricing
The engine is MIT-licensed and public — read every line before you pay for anything. The starter kit is genuinely, permanently free.
Starter kit
$0
- Starter policy file, editable, with notes
- Evidence folder and a blank draft template
- Six-question pre-publish checklist (AI Act included)
- A GitHub Actions job that gates every pull request
- Needs no API key and no network Get the kit free
Available now
The engine
$149 $39 one-off, launch price
List price $149. The launch price is $39 while the launch is running — no code needed on this link, and the list price is what it returns to.
- Claim extraction across 12 categories
- Model adjudication of every claim against your evidence
- Your policy enforced on every draft, editable
- JSON output for CI and publishing pipelines
- 120-check test suite (engine, outreach, kit)
- 30-day money-back guarantee Get it — $39 launch price
Try it on your own worst draft
Send the copy you are least confident about — paste it under the line in the email this button opens. You will get back a gate report naming every claim you cannot substantiate. Free, and no account required. If it finds nothing, you have lost nothing.
The reply you get is written by the gate itself, not by a person, and it says so. Your copy is read once to produce the report and is not stored, logged or published — which is the same rule this tool enforces on everybody else's output.