About this site

This site is not the specification. The normative text is its own document and wins over this site in case of conflict — see Documents.

Source

The site's source is github.com/algorytmi/commontime-docs, MIT licensed. The repository holds the site's content, the build script and the static server — not the specification, which is its own document.

Theme

The site uses the mkdocs-shadcn theme by @asiffer. The theme is MIT licensed:

MIT License · Copyright (c) 2026 asiffer

The licence asks for the notice to travel with the code, not to be printed on every page. So it is here and in the licence file rather than in every page's footer.

Three changes have been made to the theme, all under overrides/: the language switch (the theme renders a top-level external nav entry as a section heading with no link), this footer, and a Finnish message catalogue — the theme ships only French and Portuguese.

Fonts

The theme brings its own: Geist, Geist Mono and Inter. The site loads fonts from nowhere else.

The mark

The Common Time mark is 13 rings and 92 vertices, one path with fill-rule="evenodd". The top bar carries a culled six-ring variant, because the full mark was measured muddy at 32 px and unreadable at 16.

It receives four impulses to the bar at 118 BPM — the specification's own vector tempo — so one bar is 4 × 60000 / 118 = 2033.898 milliseconds. The phase is derived from the shared anchor, because without it every reader's mark would beat in its own phase, which on a site about getting independent clients onto one time would be its own joke.

Geometry and rhythm are identity, not protocol. The specification says nothing about the mark.

The live tick

The readout on the front page derives its own tick from a fixed anchor, with the same arithmetic a client must use: integers (BigInt) and floored division, never floating point. N1 is not a formality, and the page keeps to it itself.

prefers-reduced-motion stops both the readout and the mark.

The search index is lunr, and the Finnish build uses Finnish stemming. Each language is indexed separately, because the site is built twice — one build per language.