001What's Wrong? · Automatic QA

Checked before
anyone opens
the file.

Drop a bilingual file, or let the check run inside the editor your translators already have open. Deterministic rules settle what is certain, a model reads every segment for what they cannot see and your file comes back exactly as it arrived.

Rules first, model second XLIFF family plus spreadsheets HTML, CSV or JSON out
one segment, one findingillustrative
Source · EN

Get started with {count} free templates.

Target · TR · machine + post-edit

Ücretsiz şablonlarla hemen başlayın.

Rule
Criticalplaceholder_mismatch

{count} is in the source and missing from the target. Deterministic: same file, same finding, no model involved.

Second opinion
AI agrees

A second model read the finding and agrees. The label sits on the finding; the reviewer still decides.

Reviewer fix · made in their own CAT tool

{count} ücretsiz şablonla hemen başlayın.

EKMarked fixed by E. Kaya, reviewer v2 · 11:42
01Two layers

The rules settle what is certain,
the model reads the rest

The deterministic layer answers the questions that have a right answer, instantly and reproducibly. Then a model reads every segment, including the ones the rules passed, because the errors that cost you a client are the ones no rule can express.

  • 01

    Rules

    Tags, numbers, placeholders, URLs, empty targets, locked segments, do-not-translate terms, glossary, punctuation, spacing, expansion and locale formats. Instant and reproducible, with every rule set to error, warning, info or off in your profile.

  • 02

    AI review on every segment

    After the rules, a model reads every translated segment, including the ones they passed, for mistranslation, omission, addition, terminology and register. Your text is fenced in the prompt as data, so a segment cannot instruct the model. It runs on the provider your workspace allows: Google Gemini, OpenAI, DeepInfra for open models, or DeepSeek.

  • 03

    The second opinion

    A second model reads each rule finding and says whether it agrees, disputes it or cannot tell. The opinion is shown on the finding as a label: AI agrees, AI disputes, AI unsure. A disputed finding stays in the report and is left out of the score. You can also require a second, independent model to review each AI finding before it is kept. The reviewer decides, and the decision is recorded.

  • 04

    The report

    Findings with the rule that fired, the exact span in source and target, the second model's opinion and the suggested fix. HTML to read, CSV for a spreadsheet, JSON for your own pipeline.

02What the rules catch

The certain errors,
settled without a model

Every check below is deterministic: same file, same result, no model involved. You decide which of them can fail a delivery.

StructureTagsCount, id and order, so markup survives the round trip.
StructurePlaceholders{name}, %1$s, URLs and e-mails, plus your own patterns.
StructureNumbersFigures and amounts dropped, added or altered.
CompletionMissing targetEmpty or partially delivered segments.
CompletionLocked segmentsAnything marked do-not-translate that was edited anyway.
CompletionSame as sourceUntranslated lines, with your DNT list and brand names exempt.
TermsProtected namesBrand and product names, with the localized forms you allow.
TermsGlossaryYour term base, checked in the target.
LocaleFormatsDate order, decimal and thousands separators per locale.
LocaleURL localizationCountry code, host and the URLs that stay untouched. Runs once a profile switches it on.
SurfacePunctuation and spacingLeading and trailing spaces, double spaces, end punctuation.
SurfaceExpansionTarget-to-source length against your own thresholds.

A critical-only profile turns the cosmetic checks down to info, so a delivery fails on product names, numbers, placeholders, tags and missing targets alone.

03Profiles

One client's house style
is another's error

A profile is the rule set for a client, a locale or a content type. Clone a default, adjust it and hand it to the person who owns that account.

01

Do not translate

A list you maintain, plus detection for the tokens that behave like DNT. Those segments stop being false positives.

02

Protected terms and aliases

Brand names that must survive verbatim, plus the localized forms you have approved as correct.

03

Severity per rule

Error, warning, info or off. The same engine can be strict on one account and quiet on another.

04

Routing by folder

Files landing under a known account folder pick up that account's profile, so nobody has to choose it by hand.

05

A named owner

Each profile has someone responsible, who can edit it even when their role sits below project manager.

06

Edits stay in your workspace

Changes are stored per workspace, not written back over the shipped defaults, and the folder watcher reads the same settings.

04Where it runs

In the editor, on a folder
or in your pipeline

Nobody has to log in somewhere new. Pick whichever surface matches how the work already reaches you.

01

In the editor

A browser extension that runs the check on the job the translator has open, using their own session, on the hosts it supports: Phrase, XTM Cloud, Smartling and WorldServer.

02

Connected platforms

A Crowdin app with an editor panel and a project report. Crowdin shows the findings for each translation in its editor as a QA check result. A Smartling webhook runs on each published file, and Smartling receives each one as an issue.

03

Watched folder

Drop a package into a Dropbox folder. The report comes back beside it and the source moves to done.

04

API and CI

REST with workspace keys and outgoing webhooks, or a command-line run that exits non-zero when an error-severity finding lands, for a CI step.

05What it does not do

The limits, before
you find them yourself

A

It never edits your file

The engine reads the bilingual file and reports. Corrections are made by a person in the tool they already use, which is what keeps tags and structure intact.

  • No automatic rewrite of a master file
  • Findings are advisory: the check never blocks a workflow
  • A reviewer can override any finding, including a confirmed one
B

It says when it cannot answer

A failed or timed-out run reports segments as not checked with the reason. It does not report them as passed, because a green tick that checked nothing is worse than no tick.

  • Large project reports are capped and labelled as a partial scan
  • JSON, .properties, .strings, .xml and .resx content is out of scope today
  • The second model marks a finding unsure rather than guessing
006Try it on your own file

Send one file you
already delivered

We run it through the full check and hand back the report with every finding, its evidence and the suggested fix.