Math, as the author wrote it
A MathJax page comes back as the TeX the author typed, once per formula. A screen-reader of the rendered page gets every equation twice, with the fractions flattened.
Open source · Runs locally · MIT
The honest web reader.
Point it at a website, anywhere in the world. Every public page comes back as Markdown in the order a reader sees it, with a note of how each page was obtained — and a refusal, named, for every page it could not read.
Earth imagery: NASA
How it reads a page
Reads the page the way a person does.
Point a naive reader at a website and it returns the cookie banner as the article, a login page as the page, a 503 as text, sixty links no reader ever saw. This engine fetches twice, refuses the walls, drops what the browser hides, and writes the page in the order a reader sees it.
Discover the pages
Every crawl starts by finding out what's actually there: robots.txt, then sitemap.xml, then the links each page turns up. A queue, not a guess. On one university site that queue held 17,000 URLs, a third of them PDFs, counted but never fetched.
Fetch twice
A plain request and a real Chromium render, 1440×900. Neither alone is complete: rendering loses server markup on hydration; a plain fetch misses what scripts insert. Guessing which one a page needs was measured and dropped: the engine reads both.
Refuse the walls, drop the hidden
A wall is not a page. A login redirect, a bot challenge, a 503 are refused in the engine's own words, never passed off as the page. What the browser hides stays hidden: cookie banners, closed dialogs, links parked off the screen, display:none. On one university site that was ~60 gambling links on every page; on one government site, 2,100 words in nine closed dialogs.
Reading order, then Markdown
Every element is measured. Reading order is a recursive XY-cut over the boxes, so columns stay columns. The Markdown says whether its order was measured or assumed, how the page was fetched and what was refused. The Site Truth Report says the same for a whole site.
Build the graph
Pages, sections and the links between them become a graph: the site already is one, nothing here is LLM-inferred. Entities carry their schema.org identity across every page that mentions them.
Stay on any step as long as you like — the illustration keeps moving; it never freezes waiting for you to scroll on.
Refusals, named
A page it cannot read is a refusal in the engine’s own words, not a guess. The message names the wall.
Provenance
Every quote carries its address: the page, and the block on it. Check it without fetching again.
Each example is a real page; run the address yourself to see the same output. The Site Truth Report — what a site shows people against what it shows crawlers — runs on any site.
04 · The proof
What was measured
0.861
WCXB, official split
first of nine; 0.002 over the published rival, 0.014 run here
0.860
WCEB, 3,985 pages
second to trafilatura 0.867; 0.885 with the comment field, first
0.650
WebMainBench, tables and code
second to MinerU-HTML 0.826; column mean
0.783
DAnIEL, five languages
fourth of nine; readability 0.840 leads
Rows for this engine measured 21 September 2026 on main@928cf84 with the runners in benchmark/. Every other row says whether it was run here or is its authors’ published figure. All boards, including the ones we lose →
Refuses, and says why
A page it cannot read is a refusal, not a guess. The message names the wall.
- HTTP 5xx
A render answered with a server error is a failed side, not content.
The plain fetch stands alone and the render's status is reported beside it.
- Login redirect
redirected to a login page (…); the page requires a sign-in and nothing of it was served
- Bot challenge
Vendor named: Cloudflare, AWS WAF, DataDome, PerimeterX, Akamai, Kasada, Imperva.
the site answered with a Cloudflare bot challenge — a script a browser must run before the page is served — and no page
- Block page
it said: "You've been blocked…"
The block page's own words, quoted; never passed off as the page.
- 404 / 410
HTTP 404: page does not exist
Never rendered — a browser renders a 404 page too.
- robots.txt
The site asked; the message says what it offers instead.
The rules that applied to this client are shown with the run.
Drops what the site hides
What a browser hides from a reader stays hidden in the output, even when the HTML holds it.
- display:none, visibility:hidden
Dropped.
php.net's hidden 100-link table of contents.
- Off-screen, negative coordinates
Dropped.
Injected links a reader never sees: ~60 gambling links on every page of vtu.ac.in.
- Closed modals
Dropped.
2,100 words in nine display:none dialogs on one government site.
- Consent dialogs, share bars, “most read” rails
Left out of content.
- Alt text, screen-reader-only text
Not page text.
Kept on request; a sighted reader never saw it.
- A wall served to one fetch
Left out, and named.
The other fetch stands alone; the union does not put the wall's words back.
Refusal kinds are login · challenge · block · undeclared, from the engine’s own error types. An open cookie prompt is on the page and is judged like anything else.
Where it stands, including where it does not
| board | pages | metric | webgraph | leader | place |
|---|---|---|---|---|---|
| WCXB | 1,497 | word-level F1 | 0.861 | 0.861 | 1st of 9 (rs-trafilatura 0.859, published) |
| WCEB | 3,985 | ROUGE-LSum F1 | 0.860 | 0.867 | 2nd of 7; 1st with the comment field joined (0.885) |
| Zyte article | 181 | 4-gram shingle F1 | 0.931 | 0.970 | 11th of 35 |
| WebMainBench | 545 | edit distance, 5 columns | 0.650 | 0.826 | 2nd of 6 (column mean; tables 0.413) |
| DAnIEL (5 languages) | 1,689 | ROUGE-LSum F1, macro over languages | 0.783 | 0.840 | 4th of 9; was last before the router learned the languages |
On WCXB’s second, held-out split (511 pages) this engine scores 0.878; rs-trafilatura, run here on the same files, scores 0.891. Read every board →
05 · Run it
Point it at a site.
Runs on your machine. The only requests made are to the site you name.