How this site is verified

The rule

A claim stays on this site if it traces to a source, and the source says what the page says it says. Not what the textbook version says. Not what everyone repeats. What the paper says.

That standard is checked one paper at a time: find the paper, read what it actually claims, check the claim on every page that cites it, and attach the paper's URL to the citation. A citation with a direct link (marked "Read the paper") has been through that process. A citation without one hasn't yet. The build system counts both, so progress is measurable rather than claimed.

Why this matters here

Checking has consequences. It has already changed this site. The scaling-laws entry deleted a widely-repeated explanation after a 2024 paper refuted it. The prompt-engineering entry corrected a "first" that the cited paper's own text contradicts. The word2vec entry dropped the famous doctor-nurse analogy after the paper showing it's an artifact of how the demo excludes answers. Each of those corrections is logged, publicly, on the corrections page.

The pattern across every error found so far: the mistake is never the number and never the definition. It is the story — who did something first, what caused what. So verification here prioritises exactly those claims.

Edited, not crowdsourced

Artifipedia has one editor. That is a limitation and a position. Nothing is published by consensus, nothing is written by committee, and when something is wrong there is exactly one person responsible for fixing it. Wikis distribute trust across contributors; this site concentrates it into a method you can audit — the sources are listed, linked where verified, and the corrections are public.

Years, titles, and the arXiv problem

Citations here use the published year and venue, not the preprint year. A paper posted to arXiv in 2020 and published at ICLR in 2021 is cited as 2021. Fifteen citations on this site were silently wrong about this before the rule was enforced; all are fixed and the class of error is checked mechanically on every build.

What this doesn't guarantee

Honesty about limits is part of the method. Verification against a paper cannot catch a claim the field itself later overturns — one explanation cited here stood for four years before being refuted, and was corrected only when the refutation was found. Some claims have no correction paper to find. And roughly one paper in six has been through the full process so far; the rest carry their sources unlinked, which is the visible mark of not-yet-checked. The count on this page's build output is the honest state, and it moves.

The prerequisite graph

Every concept records what must be understood before it, as a directed relation separate from the associative "connects to" links. Direct prerequisites only: if A requires B and B requires C, C is not listed on A. Everything transitive is computed rather than authored, which keeps the relation readable and the derived numbers meaningful.

From that relation four figures are computed for each entry and shown under "Where this sits": how far into the subject it sits, how many concepts must come first, how many it opens up, and how many fields it touches. None of them is assigned by hand. They fall out of a topological walk over the graph and change automatically as the corpus grows.

Two things are built on it. Every concept page carries a small static diagram of its immediate neighbours, prerequisites above and dependents below, so the shape is visible rather than only stated. And the learning path walks the graph backwards from any concept you choose, returning the full chain of prerequisites in an order where nothing appears before the things it depends on, tie-broken so that within a level the concept unlocking the most comes first.

Two limits are worth stating. These are authored judgements about how the subject fits together, and a different editor would produce a different graph and therefore different numbers. And the relation describes the material, not the reader: someone arriving with prior knowledge is not bound by it. Nothing on this site is gated by it. It is a map of the subject, not a test.

Report an error

If something here is wrong, say so. Substantive corrections are logged publicly with what changed and why. Being corrected is the method working, not failing.