← SkillSafe / LQA Desk

Your translated batch, reviewed segment by segment

Paste the source and the target together — a TSV export, two keyed string tables, triple-pipe pairs, prefixed lines, or just alternating lines. Before anything is sent anywhere, this page aligns the segments and checks them: placeholders, numbers compared as values, markup balance, your glossary, consistency across the batch, and the target locale's own punctuation and spacing rules. Then pick a lane.

Both examples ship with a saved model run for every lane, so you can see a whole review — findings, termbase, revised targets and the locale sign-off — without signing in and without spending a credit.

nothing pasted yet
Drag a file in, or Everything is read in your browser. Nothing uploads until you run a lane.

Glossary, do-not-translate list and reviewer notes

source = target, source -> a | b for alternatives, !variant to ban one, DNT: Term to require it survives untranslated. A right-hand side of n/a, none or - means there is no agreed term yet, and nothing is reported as a miss for it.

Paste a bilingual batch to price the run.

What the free check actually does, and what it does not

It aligns your paste into segments and compares each pair mechanically. Placeholders are read across the ICU, printf, positional printf, handlebars, template-literal, Ruby and bracket forms and compared as multisets, so %2$s %1$s is recognised as a legal reordering rather than reported as two losses. Numbers are canonicalised to values before they are compared, so an English 1,250.50 and a German 1.250,50 are the same figure — and a token that is genuinely ambiguous keeps every reading it could have, because a wrong "the number changed" is worse than a missed one. Glossary terms are matched on word boundaries in Latin, Cyrillic and Greek, so log does not fire inside login, and as substrings in CJK, Thai and Devanagari, which have no word boundaries to anchor on. A glossary line whose right-hand side reads n/a means there is no agreed term, not a term spelled "n/a".

Severity is computed from the mitigating facts, not fixed per rule. A target identical to its source is informational when the source is a URL, a lone placeholder, a two-character label or a do-not-translate term, and an error when it is real prose. A missing closing tag is one finding, not two, because the derived "markup missing" is demoted and told which finding explains it. Half-width punctuation in a Chinese or Japanese target is flagged only when it directly follows an ideograph, so 1,000 and Wi-Fi, Bluetooth are left alone.

It does not judge meaning. Whether a rendering is accurate, natural or the right register is what the model lanes are for, and the checks above are handed to them as facts they must answer for — this page then counts the ones the reply skipped. Nothing here is a substitute for a native reviewer, and no lane ever edits a file: everything stays in this page until you export it.

Nothing to hand? Load the , forty strings with a swapped figure, a lost placeholder, a banned term and a broken tag, or the , where half-width punctuation, an English sentence left behind and an inconsistent term are the interesting failures. Both replay a saved run in every lane, for free.