One source, versioned translations
English remains canonical. Each translated MDX artifact records its generator, generation date, review status, and a hash of the English source so an outdated translation can identify itself.
Interactive experiment / Translation lab
Draft marks for translation
AI makes it practical to publish in more languages. It does not make every translation settled. This experiment keeps the reach and exposes the seams.
Ten translated routes now share one review surface: visible translation choices, optional learning notes, a direct path back to the English source, and an explicit AI-draft status.
Live translation surface
Read the choices, not only the copy
Toggle the annotations. Then enable English lookup and double-click any word—or select a phrase—to reveal its source.
AI-oversat kladde
Genereret på Dansk med OpenAI Codex den 2026-07-20. Kladdespor viser de valg, der blev truffet under oversættelsen. De er håndtag til gennemgang, ikke sikkerhedsscorer.
Kald dem kladdespor.Terminologisk valgOversat som: kladdespor.Det sammensatte ord understreger synlige rester af tilblivelsen frem for korrekturtegn eller et generelt kladdemærke.Alternativ: kladdemærker Ikke vandmærker i den moralske paniks betydning. De er signaler til gennemgang: synlige håndtag dér, hvor systemet traf et betydningsfuldt valg.
Under the surface
What changed to make this possible
Language-aware rendering
Routes carry BCP 47 language tags, text direction changes for Arabic, and reading-time estimates account for Chinese and Japanese scripts rather than assuming every language separates words with spaces.
Reviewable choices
Draft marks expose consequential terms, idioms, compressed concepts, and preserved metaphors. They explain the chosen rendering and keep a plausible alternative close by.
The source stays within reach
A reader can select translated text or double-click a word to reveal the corresponding English. The full translated paragraph highlights while the source popup is open.
Why this matters
Access without pretending at certainty
Language determines who can participate in online knowledge. UNESCO describes multilingual access as part of building equitable knowledge societies, and its digital-era roadmap asks language technologies to support languages that are currently underrepresented online.
But fluent output can still be unfaithful output. Recent machine-translation research treats hallucination as a serious deployment problem across both high- and low-resource language pairs. More reach therefore needs more inspectability—not a smoother illusion of certainty.
The interface does some quieter accessibility work too. Language and direction metadata help browsers, assistive technology, fonts, search, and text-to-speech process each passage appropriately. The translation is not merely swapped copy; it is marked as language.
Field note
The interface began with a problem of polish.
When generated work looks finished, reviewers need visible places to question what the system inferred, compressed, rearranged, or decided.