How we check things.
What we verify, what we refuse to publish, and what happens when we get it wrong. Written so that a reader — or a machine deciding whether to cite us — can judge the process rather than take the conclusion on trust.
What gets checked
Claims a reader would act on: throughput figures, compliance dates, capability claims, market sizing. We prioritize claims that are repeated widely and sourced thinly. A claim nobody would plan around is not worth a verdict.
Where claims come from
Most arrive from the feeds our scouts read every morning — trade press, vendor announcements, regulator pages, community threads. Readers send claims too: the form on theverdict register drops a submission into the same daily queue the scouts write to, and it meets the same gate. A submission is not a promise of a verdict. It is read, and it gets one only if it is a claim someone would plan around and there is a page to check it against — a claim with no source link cannot be checked and does not enter the queue.
No email address is asked for and none is stored, so we cannot tell a contributor what happened to their claim. What appears is the verdict, in the register, under the same rules as everything else.
How a claim becomes a verdict
The source is fetched and read. The verdict must quote the exact span it rests on, and that quote is checked against the fetched page automatically. A verdict whose evidence cannot be found in the source is discarded before anyone sees it. We do not rate claims from memory or from a summary.
If a publisher blocks automated access, we say so and publish no verdict. We do not disguise our fetcher as a browser to get around a refusal.
The five ratings
Two ratings share the question mark because they share a finding: we could not confirm the claim. Marking “No evidence” with the cross that means False would claim we disproved something we only failed to find.
A verdict is about a claim, not about whoever repeated it. Most claims we check are circulating industry assumptions rather than company statements.
What we publish and what we don't
- Every number links to its source. A claim without a working link does not ship.
- Failed checks are published. We do not quietly drop claims that did not hold.
- Public information only. Never internal data of any company.
- No affiliate links on hardware. No sponsored placement inside editorial.
Corrections
Show us a verdict is wrong and it gets corrected. Write to hello@nozzlecheck.com with the claim and what the source actually says; a correction backed by a source is acted on, and you will get an answer either way. Corrections stay visible in the page history rather than replacing what was there, because the record of having been wrong is part of the record. Currently 71 of 72 verdicts carry at least one source link; the exception is a verdict whose finding is that no supporting source exists.
Re-checks
A verdict is dated because it can go out of date. Sources are edited, an indicative regulatory timeline moves, and a figure that rested on one manufacturer page in August may be settled by an independent test in November. So verdicts get looked at again: the sources on file are re-opened, the claim is judged against what they say now, and the result is recorded whether or not the rating moves.
A re-check that changes nothing is recorded too, and it is the more common one. “Checked three times against three sources, held every time” is a different statement from “published once and never revisited”, and you should be able to tell which one you are reading. Every verdict page shows when it was last checked; where there have been re-checks, each one is listed with the page it rests on and the sentence it rests on.
A rating never moves on an opinion — ours or a machine's. It moves when a page says something, and the quote is matched against that page before the change is written, exactly as on the first pass. Where a rating has changed, the reasoning it was published with stays on the page under “As first published”. One verdict in the register has been re-checked so far.
AI assistance
Research and drafting are AI-assisted. Every claim is verified against a primary source, and a human approves everything before it publishes. We do not badge individual pages as AI-generated, because the relevant fact is not which tool typed the words. It is whether the claim is sourced and who stands behind it. We do.