What's Wrong?Translation QA

Finds what your QA tool calls fine.

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.

  • Every segment readthe model reads the whole file, not only what a rule flagged
  • Your files, as they areTrados, memoQ, Phrase, Crowdin, Smartling, XLIFF
  • Never editedwe report, your reviewer makes the fix
passed by a generic QA toolillustrative
Source · EN

The refund will eventually show on your statement.

Translation · FR

Le remboursement apparaîtra éventuellement sur votre relevé.

A generic QA tool
No issues found

Length, punctuation and terms all pass. No issues found.

What's Wrong?
Wrong meaningai_mistranslation · error

Éventuellement means possibly, not in the end. A promise became a maybe, on a refund.

Suggestion · needs human approval

Le remboursement finira par apparaître sur votre relevé.

CMMarked fixed by C. Marchand, reviewerlocked
01The report

What lands in
your inbox

Not a percentage in a dashboard. A report with the evidence attached: the rule that fired, the exact segment, the severity, what a second model thinks, the proposed fix and the name of the person who dealt with it.

Delivery
spring-campaign-tr.sdlxliff
Locale
EN to TR
Volume
4,180 words
Profile
Northwind · marketing
Checked
11 Sep 2026 · 11:42
Pre-score
87.3 / 100
218 segments read14 rule findings3 AI findingssecond opinion · agrees 9 · disputes 3 · unsure 253 penalty points over 4,180 words
SegEvidenceRuleSeverity · AI
#041Get started with {count} free templates.Ücretsiz şablonlarla hemen başlayın.Fix: {count} ücretsiz şablonla hemen başlayın.placeholder_mismatchCriticalAI agrees
#067DashboardKontrol paneliFix: Gösterge paneli · glossaryglossary_mismatchMajorAI agrees
#092Your team will love it.Ekibiniz bunu sevecektir.Fix: Ekibiniz buna bayılacak.ai_registerMajorAI finding
#153Learn moreDaha fazla bilgi edininFix: Daha fazlasılength_ratioMinorAI unsure
#186NorthwindNorthwindKept deliberately · product name on your DNT listuntranslated_same_as_sourceMinorAI disputes
EKMarked fixed · E. Kaya, TR reviewer 11:42 · fixed in Trados, not by us

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.

02The gap

Correct is not
the same as good

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.

i

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.

ii

Generic QA stops at the mechanics

Numbers and tags check out, so it ships. Nothing in the pipeline reads the sentence.

iii

Pure AI cannot sign off

A model is confident and sometimes wrong. For anything a client sees, a person has to own the call.

03What we check

Rules first, then a model
reads what the rules passed

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, and a second model gives its opinion on each rule finding before it reaches your reviewer.

  • 01

    Deterministic rules

    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.

  • 02

    AI review, on every segment

    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.

  • 03

    A second opinion, shown, not enforced

    A second model reads each rule finding and says whether it agrees, disputes it or cannot tell. 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. With second-model confirmation switched on, AI findings go through the same check.

How the rules, the profiles and the report fit together: the QA page.

04Integrations

We read your pipeline,
we do not replace it

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.

Native appCrowdinEditor panel, project report and a QA check result on each string. The in-editor check needs Crowdin Enterprise.
WebhookSmartlingA published file triggers a run and the findings are written back to the job as issues. XLIFF-family projects only.
In-editorPhraseBrowser extension, riding the translator's own session.
In-editorXTM CloudWorkbench check through the same extension.
In-editorWorldServerBrowser workbench, plus the return package.
Watched folderDropboxDrop a package in, the report lands beside it.
FilesTrados · memoQ · Wordfast.sdlxliff, .mxliff, .mqxliff, .txlf, .sdlppx, .sdlrpx
FilesXLIFF 1.2 / 2.1Format read from the content, not the extension.
FilesSheets & packages.xlsx, .xls, .csv, .tsv, .xlz, .wsxz, nested .zip
API · CIREST, webhooks, CI stepWorkspace API keys, any segment pair, outgoing webhooks. A CI step that fails the build on an error-severity finding.

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.

05How it works

Four steps, no change
to how you work

No migration, nothing to install on a server. Nobody learns a new tool. The check appears inside the editor they already have open.

01

Connect

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.

02

Configure

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.

03

Run

Rules first, then the model on every segment, then the second opinion. Report in HTML, CSV and JSON with every span highlighted in place.

04

Sign off

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.

06Security

Your copy is
not our training data

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.

01

No training on your content

Your segments are never used to train our models.

02

One workspace, no leaks

Reports, profiles and glossaries are scoped to your workspace. Red-teamed for cross-tenant leakage; anything unmapped or unknown fails closed rather than guessing.

03

Secrets encrypted at rest

Connector credentials are encrypted and never returned by the API, including to you.

04

Webhooks that fail closed

Per-workspace signature, a five-minute freshness window and per-tenant replay dedup. A bad signature never starts a run.

05

Roles enforced server-side

Five roles plus a platform operator. Hiding a page is a courtesy; the API checks again, so a hand-typed URL gets a 403.

06

Your file is never modified

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

07Pricing

Five plans,
counted in words checked

Built for in-house localisation teams who own quality in markets they cannot read, and for LSPs who need to show a client, in writing, why human review is worth paying for.

Draft pricing Prices exclude VAT
    Early accessLQA scorecardsHuman review on MQM templates you own, pre-filled from the automatic check. Switched on per workspace after a test with you.Request access
    Early accessEngine evaluationThe same text through the engines and models you are considering, scored one way and ranked against your own reference.Request access
    08Questions

    The things procurement
    asks on the second call

    01Does this replace our LQA team?

    No. It replaces the part of their week spent reading segments that were fine. The deterministic layer settles what is certain, a second model marks the rule findings it disputes so the likely false positives are cleared first, and your reviewers spend their time on the findings that need judgement, which is also where they add the most value to your client.

    02How do you stop the AI adding noise?

    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. A second model then reads every rule finding and gives its opinion: agrees, disputes or cannot tell. 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.

    03Will our content train your models?

    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.

    04Which files can you actually read?

    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.

    05What does the free test include?

    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 second model's opinion. No card, no integration, no install. If what we flag is not worth your time, that is a useful answer too.

    06What counts as a word?

    A source word in a segment that was actually checked. 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.

    07What happens when the allowance is used up?

    The rules keep running, AI review pauses and the report says so. You can add a 50k pack or change plan. Unused words roll over within an annual plan; monthly allowances reset each month.

    08Does a translator working for an LSP need a plan?

    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.

    09Launch access

    Bring one delivery you
    already shipped

    We run it, show you what your current check missed and hand you the report. One real delivery, free, no integration.