new·The score now tells you which way it movedA brain's exam only ever grows: its own material writes questions, and so does every question a real caller asked and did not get answered. The score is a percentage over that growing set, so a brain that learned more could post a smaller number — and this week three did. One of them answered two MORE questions than the week before and showed eighteen points less. Printed as a single percentage, that reads as decline to a reader and as punishment to anyone who contributes material.all news →
mozg.beta
Sign in

How a brain learns

Documentation rots.
This is built to do the opposite.

Everyone has met a wiki that was true in March. A brain is the same material with a loop around it: it is examined against what it was made for, and the questions it fails are what decides which pages get read next. Seven steps, all of them running while you read this.

Notes held147,693
Pages read12,864
Brains examined77
Catalogue average80%
  1. 01

    One link becomes every page behind it

    Point a brain at a documentation site or a repository and the crawler works out how to enumerate it — the git tree, the sitemap, or a walk of the links — then reads each page. Files, pasted text, screenshots and PDFs go in the same door. A repository's own plumbing is skipped: changelogs, licences, CI folders and agent prompts are not the product's manual.

  2. 02

    Pages are read against the goal, not summarised

    Every brain states what it is for, and extraction reads for that: the same page yields different notes for a brain about pricing and a brain about the API. Tables stay whole, code examples stay verbatim, and a page dense enough to overflow one reply is split and read in halves rather than lost. Extraction is cached, so the same page is never paid for twice.

  3. 03

    Notes are atomic, and near-copies never land twice

    A note is one fact, self-contained enough to be read alone months later by someone with none of the context. Each is chunked and embedded, and a new note close enough to an existing one is rejected at the door — the same fact worded twice is how a knowledge base starts contradicting itself. Credentials and prompt-injection language are scanned out before anything is stored.

  4. 04

    Then it sits an exam — including on what it does not have

    Control questions are written from the goal, not from the notes, and deliberately include subjects the goal implies but the material has not covered. Those failures are the point. A fifth of every exam is the opposite test: plausible questions from outside the brain's scope, which it is supposed to refuse rather than answer. Three independent judges vote on every question, so a score means the same thing twice running.

  5. 05

    The questions it failed decide what is read next

    This is the loop. Failed checks are matched against the pages of the brain's own sources; the ones that look like answers are queued and read. Those same failed questions travel inside the extraction prompt, so a re-read is never blind — it is hunting for exactly what the score says is missing. Then the brain re-sits, and the difference is visible on its page.

  6. 06

    A real search that found nothing becomes an exam question

    When any agent searches a brain and gets nothing back, that query is filed. The most frequent misses become control questions, which pulls the whole machinery — focused re-reads, page top-ups, the score — toward what people actually asked instead of only what the author imagined.

  7. 07

    Sources are re-read when they change, and nothing is deleted

    Pages are re-fetched and fingerprinted; an unchanged page costs nothing, a changed one is read again and the notes it used to produce are superseded rather than removed. Every fact keeps the chain of what replaced it and when — which is what lets you ask a brain what it used to believe.

What it never does: go and look things up on its own

A brain reads what it was given — a source you named, a file you sent, a note an agent wrote back. It does not wander off to fill a gap from a search engine, and that is a deliberate line rather than a missing feature. Everything this product claims rests on knowing exactly what a brain read: the score means something because the material is known, the author can be paid because the material is theirs, and you can check a fact because it has a page behind it. A brain that quietly scraped the open web to look complete would have none of that, and you would have no way to tell.

Inside a source you named, though, it is free: give it a documentation site and it finds every page behind the link, comes back when those pages change, and goes hunting through them whenever its exam says something is missing.

Watch it happen on a real one.

Every brain's page shows its score, the categories it fails, and what it was asked and could not answer. None of it is taken on faith.

Open the catalogue

And what happens when everyone's misses feed the same shelf →