Text on every public Dekimu page now clears an 11-pixel legibility floor, measured rather than eyeballed. A new detector joined the weekly quality job, measuring design the same way the job already measures performance and accessibility: deterministically, against fixed thresholds, with no human sampling a page by eye. Its first run found 327 instances of visual debt across the public sites, most of it text rendered below that floor, and one approval turned the finding into a pull request per site the same day.
The detector is an open-source, Apache-2.0 engine shipped as a compiled binary, pinned to an exact version and driven through the quality job's own headless browser. No LLM, no API key: it loads a rendered page and checks real computed values, such as font size against a floor, contrast between text and background, line length, padding, and whether anything is visually occluded. Only its quality category counts toward the baseline. The engine also ships a second category of opinionated style rules that contradict our own design system by design; those are counted in a single line and never acted on.
It sits next to the Core Web Vitals and WCAG 2.1 AA checks the same job already runs weekly across the public sites, from dekimu.com, InvoiceUp and miniterms to the Dekimu ID sign-in screen and the public verification page. Same pages, one more lens on each.
The first census counted 327 instances. One rule dominated: undersized functional text, meaning kickers, labels and mockup chrome rendered below 11 pixels. Some sites carried dozens of instances, others only a couple. The pattern repeated everywhere because 10-pixel type had quietly become house style across separately built sites, each one small enough on its own that nobody had measured it against a floor.
A nudge that can never be satisfied is an alarm everyone learns to ignore.
The finding went out as a single approval request (raise the floor to 11 pixels everywhere) and came back approved. Execution became one pull request per site, each on its own branch. Some of the agent briefs, as first written, would have stopped at the strings the scanner happened to sample; the rule applied instead was the whole public surface, not just what the detector flagged on a single crawl. Authenticated dashboards were left alone on purpose: the approval covered public pages only, and the shared dashboard shell's own 10-pixel tokens are a separate decision for later.
Every pull request shipped clean and merged the same day. A live rescan afterward measured undersized text at zero on every public site, and the baseline the job ratchets against was lowered to match. That baseline only moves down through a deliberate, measured change; nothing automated is permitted to rewrite it, so the floor cannot silently erode back to 10 pixels on the next redesign without the job noticing. What is left on the design side is smaller and more debatable: 11-pixel body copy sitting right at the new floor, all-caps kickers that are a house-style choice rather than a defect, and low-contrast accent text already tracked by the accessibility baseline.
This post was drafted by an AI system from Dekimu's public engineering record and published with automated checks, without per-post human editing.
← Back to blog