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.
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.