Status: Accepted
Date: 2026-08-13

Context

The site has two halves. The instrument is an application: it fetches a trajectory, decodes it in a Worker, and draws sixty frames a second. The other half is the argument, and it is a document. It states what the project claims, what each claim was measured against, and the command that produces the figure again.

Those two things want opposite treatment. The instrument's cost is justified by what it does. A page of prose that cannot be read until a bundle has been fetched, parsed and executed is paying that cost for nothing, and it is paying it in exactly the situation where a reader is least patient: arriving from a link, on a connection that is not the one the site was built on.

The client is already a Vite build published to GitHub Pages. There is no server and no framework for pages.

Decision

The method pages are static HTML under web/method/, each an entry of the same Vite build, each carrying a stylesheet and no script.

Written as HTML rather than rendered from components. React is in the build for the instrument, and pre-rendering these pages through it would mean either shipping the runtime to hydrate prose that never changes, or adding a render-to-string step whose only output is markup that could have been written. What is published is then what was written, and "this page costs no JavaScript to read" is a property of the file rather than a claim about the build.

Written rather than generated from the Markdown in docs/. The reports are the reference and are longer than a page on a site should be; the site's prose is shorter, arranged as an argument rather than as a specification, and links back to the report for the full tables. Generating one from the other would mean either publishing the reference at length or teaching a generator which paragraphs to drop. ADR-0044 made the same choice for the validation report and for the same reason.

Sharing the design system rather than resembling it. The pages import the same tokens.css, set the same three faces, and carry the same masthead at the same height with the same hairline under it. The paper palette is the instrument's inverted, not a second palette: --paper-ink and its two steps are measured against --paper the way the ink ramp is measured against --chrome. The one value that changes rather than inverts is the focus ring, which is --brass-dim here because full brass is 2.4:1 against paper and a non-text indicator has to hold three to one.

Every figure names the file it came from, in the markup:

web/tests/method/figures.test.ts opens each named file and requires the figure to be in it. A table states its source once and every figure cell in it is checked. Where the page sets a figure differently from the way the report writes it, which is usually a power of ten, it also states what the report says.

Consequences

The masthead's markup is repeated on each page rather than defined once. That is the cost of the pages being what they are, it is fifteen lines of static markup with no figure in it, and the alternative is a template step that exists to avoid copying a nav bar.

Nothing in the client imports a method page, so a page that was renamed would otherwise leave a dead link behind and still build. Every link the instrument makes into the reading half is named in web/src/method/links.ts, and a test requires each one to be a page that exists and an entry of the build.

npm run budget fails if any built method page carries a script tag. The pages have no way of acquiring one by accident today; the assertion is there because the way a site starts shipping a bundle to a document is a change to how it is built rather than a change to the document.

The reading half is five more entries in the build and one more stylesheet, which a reader of the instrument never fetches and a reader of a method page fetches instead of the bundle.