Clean, and still wrong
Tags and numbers are right. A sentence is missing, a meaning drifted, or the register is one your style guide forbids.
Tags, numbers and terms are the easy half. After the rules, a model reads every segment for what a checker cannot see: a meaning that shifted, a sentence that was dropped, a register your style guide forbids. Every finding arrives with the words, a suggested fix and a reviewer's decision, and your file is never changed.
No card
The refund will eventually show on your statement.
Le remboursement apparaîtra éventuellement sur votre relevé.
Length, punctuation and terms all pass. No issues found.
Éventuellement means possibly, not in the end. A promise became a maybe, on a refund.
Le remboursement finira par apparaître sur votre relevé.
The rules, the AI review, the second opinion and the reviewer's decision, from one bilingual file to a signed report. It works without sound.
Not a percentage in a dashboard. A report with the evidence attached: the rule that fired, the exact segment, the category and severity it maps to, whether the AI agrees or disputes, what to do about it and the name of the person who signed it.
Illustrative content. Delivered as HTML, CSV or JSON, or into Crowdin as a QA check result on the string. Your bilingual file is never modified: the reviewer makes the correction in their own tool. Read a whole report.
Tag checkers and term validators pass copy that is mechanically flawless and still wrong. That is the half your customers actually read, and nobody is grading it.
Tags and numbers are right. A sentence is missing, a meaning drifted, or the register is one your style guide forbids.
Numbers and tags check out, so it ships. Nothing in the pipeline reads the sentence.
A model is confident and sometimes wrong. For anything a client sees, a person has to own the call.
The rules settle what is certain, instantly and without a model call. Then a model reads every translated segment for what no rule can see. It also gives its opinion on each rule finding before it reaches your reviewer.
Tags, numbers, placeholders, URLs, missing target, DNT and product-name protection, same-as-source, punctuation, spacing, length ratio, locale format. Instant, reproducible, no model call.
After the rules, a model reads every translated segment, including the ones the rules passed, for mistranslation, omission, addition, terminology and register. An omission or a reversed meaning is always critical, whatever the model typed. Your segments are fenced in the prompt as data, never as instructions to the model.
The AI gives its opinion on each rule finding: it agrees, disputes it or cannot tell. A dispute on a mechanical check, such as a tag, number or placeholder, stands only when a second model from another provider also disputes it. That opinion sits on the finding as a label. A disputed finding stays in the report with its reason and is left out of the score. The reviewer decides and the decision is recorded. An AI finding carries no second opinion and no second model removes it.
Each rule has its own severity per profile. How the rules, the profiles and the report fit together: the QA page.
No migration, no engine switch, no new place for your linguists to log in. The check runs in the editor they already have open, on a watched folder or in your CI. It is advisory and never blocks a workflow.
Not listed? The REST endpoint takes any segment pair. JSON, .properties, .strings, .xml and .resx content is not supported today. We tell you that before you configure anything, not after.
No migration, nothing to install on a server. Nobody learns a new tool. The check appears inside the editor they already have open.
Add the browser extension, install the Crowdin app, point a Smartling webhook at us or drop files in a watched folder. Read access is enough.
Pick or clone a profile: DNT list, glossary, protected terms and the severity of every rule. Each profile has a named owner who can edit it.
Rules first, then the model on every segment, then the second opinion. Report in HTML, CSV and JSON with every span highlighted in place.
The reviewer fixes it in their own CAT tool and re-uploads. We propose; we never overwrite. Each finding keeps the reviewer who last changed it.
Unreleased campaigns, pricing and product names pass through this system. Every segment is treated as confidential material and every gate is enforced by the API, not just hidden in the interface.
Your segments are never used to train our models.
Reports, profiles and glossaries are scoped to your workspace. Red-teamed for cross-tenant leakage; anything unmapped or unknown fails closed rather than guessing.
Connector credentials are encrypted and never returned by the API, including to you.
Per-workspace signature, a five-minute freshness window and per-tenant replay dedup. A bad signature never starts a run.
Five roles plus a platform operator. Hiding a page is a courtesy; the API checks again, so a hand-typed URL gets a 403.
We read the bilingual file and report. The correction is made by a human, in your tool, on your terms.
DPA on Business and Enterprise · full security detail
For freelance translators checking their own work, in-house teams who own quality in markets they cannot read and LSPs who need to show a client why human review is worth paying for.
A word counts once per check, whatever runs on it. Rules never use your allowance.
No. It replaces the part of their week spent reading segments that were fine. The deterministic layer settles what is certain. The AI marks the rule findings it disputes, so the likely false positives are cleared first. Your reviewers spend their time on the findings that need judgement. That is also where they add the most value to your client.
Three ways. The model reviews one segment at a time, with the neighbouring segments as context and your segments fenced as data, so it is not free-associating over a whole file. The same model then gives its opinion on each rule finding: agrees, disputes or cannot tell. A dispute on a tag, number or placeholder finding stands only when a second model from another provider also disputes it. The opinion is shown on the finding, and a disputed finding is left out of the score but stays in the report. And the reviewer decides: every decision is recorded against the finding, so it survives a re-run.
No: your segments are never used to train our models. Your segments are processed for the run you asked for and the uploaded file is removed within 24 hours.
The XLIFF family and its relatives: .xliff and .xlf (1.2 and 2.x), .sdlxliff, .mxliff and .mqxliff, .txlf, .xlz and .mqxlz, .wsxz, .sdlppx and .sdlrpx, spreadsheets (.xlsx, .xls, .csv, .tsv) and zips, including nested ones. Format is detected from the content rather than the extension, so an odd file name is not a problem. What we cannot read today: .json, .properties, .strings, .xml and .resx. We tell you that before you configure anything.
One real delivery of your own content, fully run, with the complete report in HTML, CSV and JSON: every finding, the rule that caught it, the span and the AI's opinion on it. No card, no integration, no install. If what we flag is not worth your time, that is a useful answer too.
A source word in a segment the AI reviewed. It counts once per check, whatever runs on it, and rules never use your allowance. Premium models count 10 words per word. 100 percent, ICE and locked segments are free when you exclude them, and only when the file carries match data. The same file again within 30 days is free.
The rules keep running, AI review pauses and the report says so. Free has 20,000 AI-reviewed words a month, Solo 250,000, Team 1,000,000 and Business 5,000,000. On Team and Business you can add a 1,000,000-word pack (Team £79 / $99 / €94, Business £55 / $69 / €65) or move up a plan. Unused words roll over within an annual plan; monthly allowances reset each month.
No. A translator inside an LSP's workspace needs no plan of their own. Solo is for one person's own work: one seat and no API.
No. Prices are final.
AI review on 20,000 words a month, rules on 500,000. No card, no integration, no call. Bring one delivery you already shipped and see what your current check missed.