0 % · reading the evidence
The same probes run against the original and against the clone. A clone measured only against itself proves nothing.
The whole method in one pass — capture, localize, verify, diff, and the counter-measurement that decides it. Every number in it was read from the same runs this page replays. It is still a recording, though: the tables below are rebuilt whenever the tool is, and the film is not.
44.8s 4 subjects every check run twice exit 1 on any difference
Pick a subject. These are real recorded runs — the page replays the NDJSON
that parity diff emitted against the evidence pairs committed
in the repository.
| status | metric | source | clone |
|---|
No screenshots, no numbers typed by hand. Change the tool, rebuild, the page changes with it.
oryzo.ai is pixels tall. verify walks the
whole thing in 21 steps and stops at five marks to record what the page has
become by then — on both sides, at the same marks. Which is also the honest
limit of the walk: on a page that cannot scroll, all 21 steps measure the
same frame.
Late scroll positions are where a site reveals the assets it never mentions in its HTML. A clone that only loads what the first screen needs looks finished until somebody scrolls.
How far each subject goes, on one scale. Only oryzo.ai measures in
document pixels; the others hold everything in a canvas, so
depth turns the wheel instead and records how far it still
changes the picture. Those numbers are indicative, not exact — a camera
that has reached its limit simply stops answering.
responsive resizes the browser to 15 widths and records what the
page becomes at each one — then does the identical run against the clone.
Below: every width, the page height the source produced, the height the
clone produced, and whether they match.
The bar behind each row is that page height to scale, so the site’s own
responsive shape is visible too — including the width where its layout
switches. responsive finds that width by measuring, rather than
reading it out of a stylesheet and hoping.
Three subjects were measured and did not reach parity. These are worth more than another green tick: two mark the edge of what the method can claim, and the third is a tick this page used to show before the tool learned to refuse it.
Every remaining error was one endpoint: /_next/image?url=…&w=1920&q=75.
That is not an asset, it is a server resizing images on request.
parity reconstructs the frontend a site ships. Server behaviour is not frontend, and is not rebuilt.
Reported ✓ parity, 11 of 11 this morning. It was worthless: both sides had rendered nothing, and two blank pages match perfectly.
Two defects hid behind that tick. The clone never held the site at all —
its 1.5 MB application arrives through
import("./App3D-…js"), a relative specifier
fetch could not see, and because the markup linked the loader
absolutely the browser simply took it from the original. No 404, no
signal, nothing to notice.
Both are fixed: relative imports are followed, and same-origin absolutes are rewritten so a clone is answerable only by itself. The site itself now refuses to start under an automated browser — every asset arrives, no error is raised, and no canvas is ever created. That one is unexplained.
4 of 11 checks differed, and the off-origin report named the reason:
hagen.challenges.cloudflare.com. The site is behind bot protection.
Authentication, paywalls, DRM and bot protection are not circumvented — by design, not by accident.
Both were found the same way every result on this page was: by running the same probes against both sides and reading the difference.
An early version of the guard blocked off-origin scripts by overriding
appendChild. The JSONP library it was meant to stop used
insertBefore.
The consequence was not theoretical. A local clone posted a real address to a live production Mailchimp list. A clone that lets a form through is not a copy of a page — it is a second front door to somebody else’s service.
The guard now covers all five DOM insertion paths, plus fetch,
XHR, sendBeacon and off-origin form submits — and
answers JSONP callbacks locally, so the interface still reaches its confirmed
state without anything leaving the machine.
This is on the page because a tool that measures fidelity has no business hiding its own failures. The fix is only trustworthy if the bug is stated.
The working mode this was built for: hand over a URL, say what you need, let
the agent do the rest — because through capture and
verify it can actually see what the page does instead of
guessing. The repository ships that brief as an installable skill: eight
phases with exact commands, the traps, and the stop conditions. Install it
once, and the agent reaches for it on its own.
loading the manifest…
The paragraph above is the skill's own description — the text an
agent weighs when it decides whether this is the tool for the job — and the
list is the section headings of the file. Both are read out of
SKILL.md when the page is built, together
with the install line, which is derived from the plugin manifest rather than
typed here. The page cannot advertise a skill the repository does not ship.
An agent without plugin support gets an install prompt to paste. It installs the tool and stops there — it does not go looking for a site to rebuild.
loading INSTALL.md…
That is INSTALL.md from the repository,
read from the file at build time rather than retyped — an install prompt that
has drifted from the repo it installs is worse than none.
npm install
npx playwright install chromium
node bin/parity.mjs capture https://target.tld capture/orig
node bin/parity.mjs localize https://target.tld clone
node bin/parity.mjs fetch https://target.tld clone clone/index.html
node bin/parity.mjs serve clone 4173
node bin/parity.mjs verify https://target.tld capture/orig --shots --freeze
node bin/parity.mjs verify http://localhost:4173/ capture/clone --shots --freeze
node bin/parity.mjs diff capture/orig capture/clone
diff exits 1 when anything differs, so it drops into CI as-is.
Node standard library plus Playwright. Nothing else.
| command | arguments | what it does |
|---|
This table is read out of bin/parity.mjs when the page is built,
so it cannot drift from the CLI it documents — including the descriptions,
which appear exactly as the CLI prints them. They are in German, because the
engine still is.