MX Explained

You reached this page from the QR code in the MX Explained PDF. Nothing happens until you choose: keep a copy of the document, or look inside it.

Why use MX

Machines now read your content before people do. An AI agent answers a customer's question about you, a search system summarises your page, and a screen reader describes your document. They work from what they can parse and verify. When they can do neither, they guess, and your name is on the answer.

MX is metadata that travels with the file. It records where the file came from, what it is for, and how it may be used, in a form a machine reads directly instead of inferring. The same record works in a web page, a PDF, an image, or a data file, so your content keeps its meaning after it leaves the site that published it.

Agents then quote what you declared, not what they assumed. You also hold a dated record of what you published, which an auditor or a partner can check for themselves. What a machine needs is what assistive technology needs too, and MX builds on the web standards you already use rather than replacing them.

This PDF is a worked example. Inspect it to see the record for yourself.

Small models get it right when the file says so

The machine reading your file next may well be a small, cheap one. A decision model such as TypeSafe's Jev never writes a sentence. It answers a typed question about a file, such as "is this current?" or "who stands behind it?", with a probability across the answers it was offered. On a small local model, one decision takes about a tenth of a second and costs nothing, so it can run on every file, every time.

Fast and cheap is not the same as right. We rebuilt a Jev-style classifier on a 3B model running on a laptop and asked it those two questions about a set of PDFs. On the bare files it guessed: it judged currency correctly for two documents in ten, and named an owner for nine files in ten that name nobody. On the same PDFs with an MX packet embedded, it scored 100% on both questions, at a median of 102 ms a decision. A hosted model such as Claude also got them right, at about five seconds a decision and a cost on each one.

The labelled fields did the work, more than the plain-language summary. A small model cannot infer what a document never says, but it reads what the metadata declares. So the cheapest machines in the chain reach the right answer about your content, because you told them. The sample is small, so treat the figures as indicative. The method and every table are in I Built a Decider and Measured It, and why cheap decisions multiply is in Nobody Got Fired for Using GEO.

What inspect does

Inspect opens the MX Inspector and fetches the PDF into your browser, where it is read locally and never uploaded. You see the MX metadata embedded in the file and whether the PDF is tagged for assistive technology. You also see the fingerprint of the source it was built from and the provenance records that travel beside it, which you can follow as an evidence chain.