The errors that reach your client, caught before they ship
Every problem below is one a translator or agency told us cost them a client, a rework fee, or hours of unpaid review. Here's what happens to each in ChatsControl.
Your text stays in the EU, isn't used to train AI models, and is auto-deleted 30 days after upload by default; the free checks don't send it to any external AI - see our subprocessors and OpenRouter's data policy.
The false-positive noise that makes people switch checks off
✕98% of your QA warnings are false positives, so you stop reading them
✓Checks are locale-aware (real CLDR rules, not blunt regex) with MQM severity, so the short list is worth reading - and you can mute whole classes of known false positives yourself, per project.
✕Thousands and decimal separators are endless whack-a-mole with regex
✓Decimal and group separators are checked per locale (1,000.50 vs 1.000,50), including the OCR split-decimal case, so you stop hand-writing rules.
✕Section numbering like 3.2.1 is read as a changed number
✓The numeric check explicitly skips section and clause numbers, so document structure never fires a false accuracy error.
✕A number inside a product name like “Gizmo 1000” is flagged as a changed figure
✓We pair characters exactly so a product-name digit doesn't fire - and we deliberately never add a heuristic that could mute a real change to a dosage or an amount in a CRITICAL numeric check.
✕“Source is same as target” fires on the client's brand, over and over
✓A brand kept verbatim across the whole file collapses into one finding with its occurrence count, instead of the same warning N times.
✕A borrowed termbase full of words like “home”, “in”, “of” - and typos
✓Glossary hygiene flags stop-words, too-short and over-frequent entries, duplicates and conflicts; a separate check catches likely typos in glossary terms.
✕Because of the noise, the REAL errors slip through
✓That's the whole point: less noise means real defects don't drown, and an AI semantic pass catches the meaning errors deterministic checks can't see.
What ships to the client broken - the expensive category
✕An untranslated source segment shipped straight to the client
✓Untranslated and empty segments are flagged before delivery - the one-click miss that reaches the client in every CAT tool.
✕“X is not Y” came back as “X is Y” and passed proofreading
✓A semantic pass catches meaning inversions, swapped subjects and reversed direction - the fluent-but-wrong MT output that sails through a human read.
✕Text inside an image never got translated - 44 words, a €780 layout bill
✓We OCR image-only PDFs and photos your CAT tool can't import at all; catching text baked into embedded images inside an otherwise-editable file is next on the roadmap.
✕Pages after a blank page were skipped entirely, and the client found it
✓The scan pipeline processes every page; an explicit page/block completeness signal (“content possibly missed on page N”) is on the roadmap, built on data the OCR already emits.
✕A 100% TM match dragged in another client's product name - £315 off the corrector
✓Glossary and forbidden-term checks catch a wrong brand or banned term in the target; a dedicated TM-contamination signal is planned.
✕A number written as a word (“2” → “twice”) is flagged as a missing figure
✓The numeric check already suppresses many of these; broad spell-out coverage is being verified on real corpora rather than promised blind.
Scans, OCR and layout - what CAT tools can't do
✕An image-PDF can't be imported, but everyone assumes it can
✓We OCR image-PDFs and photos - exactly the files CAT tools reject - and turn them into translatable, layout-restored documents.
✕Rebuilding the layout of a scan eats unpaid hours
✓We rebuild the layout 1:1 automatically, painting the translation back onto the original scan - the restoration work that usually isn't billable.
✕Hyperlinks broke when the file went from the CAT tool back to Word
✓Our PDF-to-DOCX conversion preserves hyperlinks, so they don't quietly disappear on the round trip.
✕Zero-width characters wrecked your match rates - invisible, and nobody checks for them
✓We flag zero-width and bidi characters (U+200B, BOM, soft hyphen, RLO) that most QA tools never look at, before they corrupt your TM leverage.
✕A password-protected PDF lands Friday evening and quietly breaks the import
✓We detect a password-protected PDF the moment you upload it and say so plainly, instead of handing back an empty result you notice on Monday.
✕Nobody checks the OCR - indices, superscripts, Greek letters in formulas
✓Overlay keeps the original scan visible for verification; a low-confidence OCR flag (per-token score, Greek-in-formula) is on the roadmap on data we already capture.
✕Numbers, sums and dates aren't in the word count, so that work is unpaid
✓Our calculator counts by characters (numbers, dates and sums included), and OCR gives a real count straight off a scan.
Machine-translation quirks
✕The same term comes back translated four different ways in one document
✓An AI consistency pass flags a single source term rendered several ways across the document - no glossary required - so it reads consistently.
✕Terminology drift across a document is a reputation risk
✓Glossary adherence and forbidden-term checks plus segment- and term-level consistency keep terminology aligned document-wide.
✕The MT output is smooth but wrong, and it passes review
✓The semantic pass targets exactly this - meaning that reads fine but says the wrong thing - rather than only surface fluency.
✕Terminology checks don't tell a noun from a verb, so a valid form reads as a missing term
✓The glossary is morphology-aware for Russian and Ukrainian, so one entry covers inflected forms instead of firing on every variant.
✕The AI knows its typical mistakes but can't catch them itself
✓The semantic pass flags those recurring MT failure modes so a human sees them, instead of trusting the model to police itself.
See it on your own file
Upload a finished translation and read the findings in under a minute. No account needed.
Check a documentQuestions translators ask
Is this a replacement for my CAT tool?
No. It's the QA and scan layer around whatever CAT tool you already use - run a finished translation through it before delivery to catch what slipped past.
How is this different from Verifika or Xbench?
Those are strong deterministic QA tools. We add locale-aware suppression to cut the false positives that make people switch checks off, plus an AI semantic and term-consistency layer and a 1:1 scan-restoration pipeline they don't have.
Which languages does it cover?
The QA checks are locale-aware for the major European and CIS languages; the scan and OCR pipeline works on any script.
Do I have to send the client's document to your servers?
You upload a document to run the checks; you control how long anything is kept in your settings. For a quick look, the QA validator runs without an account.
Is there a free version?
Yes - the QA validator and the price calculator both run for free without an account.