Real screens, one real patent, in five scenes. Every scene is spoken; mute a pane if you would rather read.
Four independent AI vendors argue the finding against its own source text. What survives becomes a Brief, and the Brief becomes one signed PDF with an independent trusted timestamp — verifiable by whoever you hand it to, without taking our word for anything.
Every screen below is from one recorded session on 19 October 2026. Nothing is mocked up.
IPVerifier is an add-in, not another portal to log into. Every analysis Reporter finishes appears in the pane on its own; one click flows the latest version into the open Word document, bound and jump-ready. Nothing to browse, download, or choose.
Enter the subject patent. Each lane card lights up when Reporter has produced that output.
Source text is fetched first. Then one model argues the finding is supported, a second argues it is not, a third judges, and a fourth vendor gives a second opinion. Each runs against the source text at temperature zero, and each reports its own elapsed time.
The real patent text, fetched before any model speaks. The council reads the document, not a summary of it.
Argues the finding is not supported by the source.
Argues that it is.
Reconciles the two against the text and grades each statement green, amber or red.
An independent review from a fourth vendor that took no part in the argument.
WO2007137244A1, seven sections, focused on the evidence. Thirty to ninety seconds.
On one recorded public sample the verdict came back do not ship, because the analysis had built an argument the source would not carry. That outcome is the point of running it: a review that can only agree is not a review. On this run it came back verified with caveats — and the caveats are the interesting part.
Every statement in the analysis, graded against the text it cites — with the argument on both sides kept.
Every finding is one click from its sentence in the document.
A raw machine translation of a foreign patent is searchable but not citable. In IPVerifier a translation can be audited for errors, reviewed block by block against the original script, and signed off by your own expert before it carries any weight. The original text sits in Word; the review sits in the pane; each block in one is linked to its counterpart in the other.
CN101641101A — the original script in Word, 86 aligned blocks in the pane, linked.
The findings you keep become one self-contained PDF, signed by the IPVectors™ Attestation Service, countersigned with an independent RFC 3161 trusted timestamp, and hashed into a tamper-evident chain. Open it in Adobe Reader and the Signature Panel shows the signer, the trusted time, and whether a single byte has changed since.
The timestamp authority is independent of IPVectors, which is the whole point. Our word is not what makes the file verifiable.
One model, asked a hard question, is confidently wrong in ways its own output will not reveal. IPVerifier puts four models from four vendors on the same question and lets them cross-examine each other. Agreement across independently trained models is evidence. Disagreement is a flag — and it is recorded, not averaged away.
Keys are stored on your computer only, encrypted before they are sent, never in our database and never in our logs. On your own keys your contract with each vendor governs; the Security tab shows what each vendor commits to, and when it was last checked.
