First run on sign-up · No LLM in the loop · Every check is reproducible
Language models invent patent numbers, quote passages that were never written, and produce measured values nobody measured. Paste the text and we run deterministic checks against real patent databases — then hand you a list of items to verify, each with the exact wording it came from.
First run included on sign-up · 1 credit per run after · Results in seconds
Paste patent text and we check the checkable parts against real data. No language model involved.
Meant for drafts you or an AI wrote and have not filed yet. Published patents work too — these particular signals just tend to show up less often in them.
The failures that matter all look perfectly plausible on the page
A fabricated publication number has the right country code, the right length and a believable year. Nothing about it reads as wrong — the only way to know is to look it up.
The cited patent exists, so the citation passes a glance. But the sentence in quotation marks may appear nowhere in that document.
"Compressive strength measured at 52.3 MPa", "experiments show a 30% improvement" — precise, confident, and often generated rather than observed. Fabricated experimental data is exactly what filing rules prohibit.
A to-verify list you can act on — not a score, not a pass mark
Every number in the text is looked up. Found ones come back with their real title and dates so you can confirm them yourself; ones we could not find in our covered libraries are flagged for you to check. Lookup failures are reported as failures, never as "does not exist".
When a quotation is clearly attributed to a specific patent, we fetch that document and try to locate the passage verbatim. If it cannot be found, you get told exactly that — and which text layer we searched.
Claimed measurements, statistics, test-standard codes and precise improvement percentages are listed for verification. Dates asserted about a cited patent are compared against the real record, and statute citations are surfaced for a currency check.
Claims, description, background, an office-action response — any patent text a model helped write. You can optionally paste your own source material so figures that trace back to it get de-emphasised.
Extraction is done by rules, verification by real database lookups, and the conclusions are rendered by code. No language model takes part — which is why the same input always produces the same result.
Each item carries the exact wording it came from, so you can go back to the draft and check it in seconds. The report also states plainly which checks ran, which did not, and how far our coverage reaches.
We check specific, checkable things: whether a cited number exists in the libraries we cover, whether a quoted passage can be located verbatim, which figures assert measurements, which statutes need a currency check. We list AI-typical phrasings sentence by sentence, but **we do not conclude that AI wrote it** — and a low reading does not mean it was not AI-written. We also do not assess whether an application complies with filing regulations — that depends on whether real research and development took place, which no software can determine. Every result is a to-verify list for a human to work through.
Sign-up includes 1 run; after that each run costs 1 credit. New users also get 25 bonus credits. It runs no language model, so it costs very little to run.
No, and we deliberately never phrase it that way. It means the number did not turn up in the libraries we cover. Those libraries are offline snapshots, so a very recent publication may not be indexed yet, and some jurisdictions are outside our coverage entirely. It is a strong signal worth checking — not a conclusion.
No. A report with no items means these four checks found nothing within the coverage stated in the report. It says nothing about novelty, claim scope, drafting quality, or whether the underlying research actually happened. That is why every report shows what was checked and what was not.
No — no software can do that reliably. What we can do is list, sentence by sentence, the phrasings that AI reaches for and that never appear in the real granted-patent passages we calibrate against. That is wording evidence, not a verdict on authorship: real attorneys write that way sometimes, and AI text that avoids those words will not show up here either.
Because the check cannot tell where a figure came from — only that it asserts a measurement. If you paste your own source material alongside the text, figures that also appear there are de-emphasised. They are still listed, because the same value appearing somewhere in a source does not prove it supports that particular statement.
Results are cached briefly so that re-submitting the same text does not re-query the databases. Your input is used only to run this check.
Disclaimer
This tool performs rule-based checks and database lookups on the text you provide. It is not legal advice, not a patentability opinion, and not an assessment of whether an application complies with filing regulations. Results are items for a human to verify. Decisions on whether and how to file should rest on the formal opinion of a qualified patent attorney or agent.