[drmowinckels]

Why

People arrive unsure whether we can help them, and the copy was a large part of
why. The site described itself in terms a reader only understands if they
already know what we do — "data engineering", "FAIR principles",
"containerization", "integrated solutions", "registry data handling 55%".

That was the starting point. The branch grew from there into the structure
around the copy, the team pages, and a set of things the audit at the end
turned up underneath.

Copy and structure

Services are problem-led, and there are five of them. Each leads with the
problem the reader arrived with, lists the situations that produce it, and says
what they get. The category name stays as a small secondary label, so someone
who does know the term can still find it.

Before   Data Engineering
         Collecting and managing data and keeping track of its lineage
         is a critical aspect of scientific research...

After    You can't work with your data yet.
         - Sixty spreadsheets with inconsistent column names.
         - A Nettskjema already running, with no plan for getting the
           data out of it.
         → one documented dataset you can actually use, and the script
           that produces it, so it reruns when more data arrives.

Seven became five, in the order people meet them: data management and sharing →
data capture and engineering → processing, analysis and computing →
applications and databases → workshops and training.

"What we run and support" is a second layer, not the spine. Barbara's four
pillars are real, but they are specific tools rather than services, so they sit
in their own block — GOAT, Nettskjema, MRI, behavioural experiments — on both
the landing page and /services/. The repetition is deliberate: it is for
anyone arriving looking for one of those by name.

Two claims came out. We do not advertise MR scanning, and we do not
advertise Bevex, which is not in beta until spring 2027. The experiments card
says what is true today: JavaScript tasks that run anywhere, the tachisto
builder, and the platforms you can actually upload them to.

/services/ is two sections. It had grown to five — services, tools, "what
working with us is like", testimonials and a closing CTA. That was the
"scattered" note applied to the page rather than to the services block. 36.9K
to 17.8K.

Skills, pricing, hero, contact. The percentage bars are gone; "registry
data handling 55%" was not sourced from anything and not something a visitor
can act on. Pricing explains that the rate depends on how close you are to the
department and that computing time is applied for separately. Contact says
explicitly that you do not need to know the technical name for your problem.

The team

Bios written from the employee pages, with summary fields so cards show a
line rather than the whole bio, and task tags at one consistent level of
granularity. Person pages lay the photo beside the text instead of under it.
Silje Forsland added; José-Luis's 2024 portrait and his statistics focus;
Markus's MRI protocol work. Employee pages are linked through an icon drawn
from the colon in the UiO lockup. Photographers are credited where we know who
they are.

Publications moved from ORCID to OpenAlex, which returns the full list in one
request and does not need a personal address in the query.

Open source

The portfolio shows everything, not just the four entries that happened to have
a logo, with project marks back on the cards — grey until you point at them.
jsPsych Studio is tachisto throughout, with its new wordmark; ggseg is now
the ggsegverse mark and an entry that describes what it is today.

Accessibility

• Links in prose are underlined by default. Teal on body text is 1.36:1,
  where 1.4.1 wants 3:1 for colour alone. The fix existed four times over and
  had already missed .person__bio, .contact-reassure and
  .help-card__short.
• Six labels that only a screen reader hears were hardcoded English on the
  Norwegian site — "Toggle navigation", "Back to top", " on github.com".
• Three copies of an aria-label built by splitting an icon class on hyphens,
  which announced "github" rather than naming the person.
• The reveal animation can no longer hide the page. It set opacity: 0 and
  waited for a script that might not arrive; the flag that hides content is now
  set by that script itself, inlined in <head>, and removed again if anything
  throws.
• Heading order, focus ring contrast, card link destinations, and a language
  switcher that is links rather than a navigating <select>.

Infrastructure

The canonical host was wrong. baseURL said capro.dev, CNAME said
www.capro.dev, and the apex redirected to www — so every canonical, og:url,
hreflang and sitemap <loc> named a URL that 301s. The apex is canonical now
and the CNAME matches. DNS has since moved to Cloudflare; both names are
proxied, the edge certificate covers them, and /.well-known/acme-challenge/
still reaches GitHub so its certificate can renew.

The deploy was 12M, of which 6.3M was images nothing linked to. Team,
testimonial and news bundles published their originals next to the derivatives
templates actually use. With the icon fonts also gone it is 4.8M.

Both icon fonts are gone. 29 glyphs were costing 737K of font files and 93K
of CSS, including two Font Awesome faces nothing referenced and three Themify
formats no browser asks for. Icons are inline SVG now, synced from the Font
Awesome package by npm run sync:icons. This also removes the
media="print" onload trick, which was a workaround for a stylesheet too heavy
to block on.

Tailwind was scanning assets/js/vendor, so minified Alpine and Leaflet
contributed utility candidates, and the bundle depended on build order — a full
npm run build and a bare npm run build:css produced different output from
the same sources. Render-blocking CSS is 59,831 → 41,088 bytes across the branch.

Also: robots.txt (a 404 in production), hreflang="x-default", Dependabot for
both npm projects and the actions, npm ci instead of npm install so CI
cannot resolve versions you do not have, and the deletion of themes/capro-hugo
— 172 files of jQuery, Bootstrap and Slick that nothing referenced.

Still to do before merge

• Bios need sign-off from all five people.
• Norwegian needs a pass. DeepL from the English, which is the source of
  truth. content/offerings/_index.no.md is two revisions behind its
  English twin — its experiments card has neither the platform list nor the
  tachisto mention. The portfolio category chips are also my Norwegian.
• Photo credits. Only José-Luis's is known (Erik Engblad).
• GOAT wording. "Built through UiO's HUB-node initiative" is the one
  phrase an outside reader will not parse. Keep only if it earns its place
  on credibility.
• AAAA records were not needed in the end — Cloudflare's edge provides
  IPv6 — but worth confirming nothing else in the org hardcodes
  www.capro.dev, since those now take a redirect.

Notes for review

• content/feature and content/cta are no longer rendered anywhere. Left in
  place rather than deleted, pending a decision.
• --color-tertiary (gold) is still defined and still unused.
• static/webfonts/leaflet ships five PNGs that Leaflet's stylesheet looks for
  at a different path. Nothing requests them — the map uses an inline-SVG
  marker and adds no layers control — so they are dead weight rather than a
  live 404.

Test plan

CI covers formatting, the content accessibility lint, a strict build, the CSS
bundle freshness check, the link check and the axe audit. Beyond that, verified
by hand in Chrome:

• Icons. 206 visible inline icons across both languages at 1280px and
  390px, all sanely sized. Two bugs only visible there: an auto-width SVG
  stretches to its container under display: block, and again as a flex item.
• Reveal. Five paths — normal, JS off, no IntersectionObserver,
  constructor throwing, observe() throwing. The last four leave all sections
  at opacity: 1; the first animates and reveals on scroll.
• Links. 167 links in the opt-out contexts render undecorated; an anchor
  injected into each of the ten prose containers comes out underlined.
• Pruned images. All 2300 local references in the built output resolve;
  none of the removed files was reachable from any page.

🤖 Generated with Claude Code

[github-actions]

✅ Preview deployed at https://25--capro-uio.netlify.app
Please check it looks ok before merging.

[github-actions]

✅ Accessibility: no issues

axe-core over 19 routes × 2 viewports (one page per layout), plus 69 manual checks.

All 69 manual checks passed.

_{Automated tooling catches roughly a third of accessibility problems. A clean run is a floor, not a ceiling.}