Skip to content
Meltemi docs

Ask the docs

Ask a question in your own words: the answer is written from the portal's pages only, and names the pages it used.

0 of 500 characters

Your question is sent to Cloudflare Workers AI to be answered and is not stored. Answers can be wrong: check the sources. Privacy notice

Answers run on Cloudflare's free allowance, about four hundred a day for everyone.

Español

0032. Reference machine and quality gates

Status: accepted (acceptance-and-quality change, 11 October 2026)

Context

Section titled “Context”

The proposal’s release criteria are about the product as a whole: an acceptance corpus in four profiles, zero cut or overlapping labels, visual tests with pinned resources, deterministic geometry and frames, fluid interaction measured on a reference machine, a first export in under ten minutes, and exports verified where they are used. It left the reference machine and the numeric targets as pending decisions. Every package had its own tests, but nothing measured these criteria together, and a first audit of the drawings found two faults no check saw: a connector of an assisted C4 view ran through a box with its label on top, and a connector entering a group ran over the group’s title.

Decision

Section titled “Decision”

  • The reference machine is the one the product is built on: Intel Core i9-11980HK (8 cores, 16 threads), 64 GB, Windows 11, Node 24, Edge 155, Chrome 154, Firefox 157 and FFmpeg 6.1. Every recorded figure names the machine and versions it ran with.

  • Interaction targets, at the 95th percentile on the reference machine, for views of up to 30 elements in each profile:

    • the analysis a file gets after an edit (compile, layout and diagnostics inside its project): under 500 ms;
    • a canvas drag from the message to the new drawing (plan, edit, analysis, render): under 100 ms.

    Tests assert them with a factor of four, so that shared CI machines do not fail by chance; pnpm bench --write records the exact figures in docs/performance.md.

  • First export: from init to exported SVG, PNG, PDF and MP4, the machine’s part is under 60 s for every template, which leaves the proposal’s ten minutes to the person reading and changing the template. The playground path is timed too.

  • Labels: zero cut or overlapping labels over every corpus, template and example. The layout audits each drawing it makes and reports layout.label-cut and layout.label-overlap; validate reports layout diagnostics, so validate --strict in CI fails on them.

  • Visual references are PNGs at scale 1, drawn by the installed Edge or Chrome with the bundled fonts and the pinned ELK. A pixel differs when a channel differs by more than 16 of 255; a drawing passes with at most 0.1 % differing pixels and no differing run longer than 2 pixels. The manifest records the browser version; references change only with --update.

  • Determinism is checked by content: SVG and HTML byte for byte, PNG pixel for pixel, video frame by frame with FFmpeg’s frame hashes, so container metadata never counts.

  • Exports are verified where they are used: automatically in Edge, Chrome and Firefox (Firefox through WebDriver BiDi, with the browser already installed), including the playground’s own exports; by hand in PowerPoint, Keynote and Google Slides with a kit of files and a checklist whose results table the owner fills in.

  • One command runs every criterion: pnpm acceptance writes docs/quality/results.md.

Alternatives considered

Section titled “Alternatives considered”
  • A cloud CI runner as the reference. Reproducible to anyone, but shared runners vary by tens of percent between runs, and the proposal speaks of the machine people work on.
  • Frame rates as the interaction target. The canvas draws with the browser; what the person waits for is the round trip from a drag to the new drawing, which is what is measured.
  • Strict pixel equality. Anti-aliasing differs between GPU paths of the same browser version; the tolerance absorbs it, and the run limit still catches a moved label or line.
  • Comparing video files byte for byte. WebM muxers write time-dependent metadata; frame hashes compare what is seen.
  • Driving slide applications with UI automation. Fragile, licence-bound and different on every version; a person with the applications checks them with the kit.
  • Downloading Playwright’s own browsers. Not approved, and not what people use; the installed Edge, Chrome and Firefox are driven instead.

Consequences

Section titled “Consequences”
  • The layout keeps labels clear by itself: relation labels slide to a clear stop, straight connectors pass under group titles in pieces, and default connectors go around boxes in assisted and manual layouts; layout.connector-crossing reports what still crosses.
  • Exported players and catalogs carry an empty inline icon, because Firefox asked the server for /favicon.ico and logged the policy’s refusal as an error.
  • packages/acceptance is a private package of tests and runners; its figures are only meaningful with the machine line next to them.
  • A new browser version may change pixels within the tolerance; a larger change is reviewed and recorded deliberately.