Confident AI that invents numbers is worse than no AI at all — especially in a document you are about to send to someone who matters. Each of these is checkable in use, which is the only thing that makes the mark worth anything.
A hard number with nothing behind it — a market size, a percentage, a "research shows" — is caught before it reaches you and returned to the model to hedge, cite, or retract. Survive a second pass and you see it flagged.
Gates come back weak or dead with a reason. Flattery here would cost you months, so you won't get any.
Your idea and your API key stay in your browser and go straight to your provider. No account, nothing on our servers. We do count page visits with Google Analytics — never your idea, never your key.
Every claim carries its support — proved, asserted, or estimated. The marks live inside the document and survive export, print, and re-import.
A governance claim nobody can audit is decoration. So here is the mechanism, exactly, and the way you would check it yourself.
Every response is scanned before it reaches you. A specific factual claim — a percentage, a market size in millions, "research shows…", "according to…", "it is proven that…" — asserted with no hedging and no stated basis is a hard violation.
How to checkAsk it for a market size. If a hard number comes back, it should be hedged or attributed. A bare "the market is 4.2 million firms" is the exact shape the check exists to stop.
A caught response is not edited or annotated — it goes back to the model with the offending sentence quoted and an instruction to hedge it, cite it, or retract it, then return the whole answer again. You see the rewrite, not the original.
How to checkWatch the status line while a forge runs. When the check fires, it says so before the second pass.
If the rewrite still asserts the claim, the tool stops trying and surfaces it: the response arrives with the unverified claim named at the top. Quietly passing it on after two attempts would be the more comfortable option and the dishonest one.
How to checkAn uncorrected claim appears as a warning line above the result, quoting the sentence.
The fastest way for a trust mark to become worthless is to claim more ground than it holds. So, plainly:
Not a certification. Nobody audited this and nobody accredited it. It is our own protocol, applied to our own product. Treat it as a promise we can be held to, not a seal somebody issued us.
Not a guarantee the answers are right. It governs how uncertainty is handled, not whether the model is correct. It reduces the risk of confident invention; it does not remove it. Everything it produces remains yours to review before you rely on it.
Not the reason your data stays private. Nothing reaches our servers because the tool is a static page running on your own API key — that is architecture, not ethics. Worth having, but it is not INTEGRA.
Not a stamp on your document. Your Sell-Kit carries a disclosure, never our branding. Marking individual outputs as approved would be exactly the kind of unearned authority this protocol exists to refuse.
Ten minutes, no product required
Free, in your browser, on your own key. If the idea is bad, you'll know before lunch.