Ask a translator and you get the same four stages back: analysis, transfer, restructuring, testing. The names come from Eugene Nida, who set them down in 1969, and they have outlived every tool built since.
Most people do two of them. They read the sentence, they write the new one, they call it done. The two that get cut are the two that decide whether the result is any good.
Here is what each stage is for.
Stage one: analysis, work out what it says
Before a single word goes into the new language, you take the old one apart. Not the vocabulary. The meaning. Who is speaking, to whom, how formally, and what the sentence is actually doing underneath the words, whether that is stating a fact or softening a refusal.
Nida was blunt about this. Most translation errors are analysis errors, made by a translator who moved to the writing too fast on a reading that was almost right.
Take the Russian toska. A quick reading says sadness and you are already off. A real one finds a restless ache with no cause you can name, sitting somewhere between longing, boredom, and grief. Get that wrong at stage one and every later stage just polishes the wrong sentence.
Stage two: transfer, carry the meaning across
Now the meaning moves. This is the only stage that lives in two languages at once, and it happens in the translator's head, not on the page. You hold the sense of the original and reach for the nearest thing the other language keeps.
Every transfer is a trade. The skill is knowing what you are giving up.
Sometimes there is a clean match. Often there is not, and you have three moves: add, drop, convert. The Portuguese cafuné is running your fingers through someone's hair, tenderly. English never bothered to keep a word for it, so transfer stretches it into a phrase. You lose the compactness to keep the meaning, and you make that call on purpose.
Stage three: restructuring, make it sound like the language
Transfer hands you the right meaning in roughly the right words. Restructuring turns that into a sentence a person would actually write.
Word order shifts. A noun turns into a verb. A long German chain breaks into two English sentences, because English does not stack that way. Torschlusspanik is the panic that the door is closing on your life, that the time to do the thing is running out. In English it becomes a clause instead of a word, and the sentence around it has to be rebuilt to carry it.
Register lands here too. Formal, casual, warm, clinical. A sentence can be accurate and still be dressed for the wrong room.
Stage four: testing, find out if it landed
The last stage is the one that gets dropped. You put the translation in front of a reader who never saw the original, and you watch what they take away.
The question is not whether it is correct. It is whether it does to them what the original did to its own readers, whether they catch the joke and hear the politeness that was sitting in the original.
Nida had readers say it back to him. Publishers call it review, editing, proofreading. The name changes, the job does not: a second pair of eyes checking the output against reality instead of against the source.
The other four stages you will see
Search the phrase and you will also meet the industry version: translation, editing, proofreading, quality assurance. TEP plus a final check.
Look closer and it is the same shape wearing a job title. Analysis and transfer collapse into "translation", restructuring becomes "editing", testing splits across proofreading and QA. One model describes what happens in a translator's head. The other describes who signs off. Both put a check at the end, for the same reason.
Where machines stop
Machine translation is genuinely good at stages two and three. It transfers fast, and modern models restructure well enough that the output reads smooth.
Stage one it fakes, and stage four it skips entirely. It does not know the register of the room you are writing into, and it never asks whether a real person would say the sentence it just produced. It sounds equally sure of itself either way, having checked nothing.
That last gap is the expensive one. A sentence can be grammatical, natural sounding, and still be a phrasing no native speaker has ever used.
How Linguin runs the last stage
Linguin runs the testing stage instead of cutting it.
Turn on live mode and translation runs in two passes. The first drafts the sentence. The second takes the draft's own phrasing and searches how people write that language right now. Wording that turns up in real use stays. Wording that does not gets rewritten before it reaches you. It is Nida's read-it-back-to-a-native, run against a live corpus. Builders get the same pipeline through the API.
Stage one gets the same treatment. Look up a word and you get the register it carries and the moments people actually reach for it, in your own language across 120+ languages. The Italian sprezzatura is not "nonchalance". It is studied effortlessness, hard work staged to look easy, and that difference is the entire word.
The two stages nobody bills for
Analysis and testing are the unpaid ends of the job, and they are where the quality actually comes from. Everything in the middle is mechanics.
So when you are checking a translation, work from the ends in. Ask what the original was really doing, then ask whether anyone would say what came back. Start with the first question.