Verifika alternative for online translation QA

What's Wrong is an online translation QA tool: the form checks you know from Verifika, then a reading for meaning. On 10 of our delivered jobs it found 22 real errors Verifika missed and skipped 84% of its false positives. Verifika corrects in place; What's Wrong never touches your original and returns a corrected copy of the fixes you accept.

Verifika was part of our delivery routine for years; our own archive holds Verifika reports on Spanish, English and German jobs into Turkish and Arabic from 2018 to 2022. It is fast, it is thorough on form and the reviewer can fix a finding without leaving it. Linguists love Verifika's interface and its apply-fix. We learned from that for years and built the same care into What's Wrong.

What it cannot do is tell you that "safest" became "fastest" or that a washbasin tap became a bath tap. Those are the errors that reach a client, and they are the ones this page is about.

Verifika and What's Wrong side by side#

What we measured on our own files#

We took 10 jobs we had delivered, Turkish and Arabic, that still had the Verifika report next to them, most of them with the reviewer's own false-positive labels on each row. Then we ran the same files, as they were when Verifika saw them, through What's Wrong on production with the AI review on. About 18,600 source words in all. On four of the files the AI review read a sample of the segments, so the counts there are a floor.

• 22 real errors Verifika never flagged. A reviewer would have changed every one of them: a meaning reversed, a typo that turned a technical term into nonsense, a misspelt scientific term. Ten more sat inside segments Verifika had flagged for something unrelated, a number or a quotation mark (a washbasin tap called a bath tap in the terms of a prize draw is one), so the reviewer had looked at the segment and moved on.
• In one website job, at least eleven of those errors reached the client. We checked the delivered file. The mechanical findings had been cleared; the wrong words were still there.
• 84% of Verifika's false positives did not appear at all. Of the rows the reviewers had marked false positive or ignore in Verifika, more than eight in ten were simply not in the What's Wrong report. More than half of the rest came with the AI disputes label, which keeps them visible and takes them out of the way.
• Where the two overlap, they agree. Every real spelling or name error in Verifika's lists, What's Wrong found too. Verifika is the more thorough of the two on quotation marks: on one file it caught a quotation style our default profile did not.

That is the honest shape of it. Verifika is good at form and can fix it in place. What's Wrong is good at form and reads for meaning. Updated November 2026, from our own delivered files; the raw reports are not public because they hold client text.

Three errors Verifika did not flag#

Names and brands are changed; the error is kept word for word.

1. The safest became the fastest (English to Turkish, shipped)

• Source: Designed to provide the safest operator experience, even when working with mercury
• Target: Cıvayla çalışırken bile en hızlı operatör deneyimi sağlamak için tasarlanmıştır
• Verifika: not flagged.
• What's Wrong: "The word 'safest' was mistranslated as 'en hızlı' (fastest) instead of 'en güvenli' (safest)." Suggested fix: Cıvayla çalışırken bile en güvenli operatör deneyimi sağlamak için tasarlanmıştır

2. A washbasin tap that became a bath tap (English to Turkish, prize-draw terms)

• Source: Each set contains one Linea L-size washbasin faucet, one head shower set and one towel ring holder, all in the same Aquor colour.
• Target: Her bir set, hepsi aynı Aquor renginde olmak üzere bir adet L boyutunda Linea küvet musluğu, bir adet duş başlığı seti ve bir adet halka tipi havlu askısı içerir.
• Verifika: flagged the segment for its numbers; the reviewer marked it ignore.
• What's Wrong: "The source specifies a 'washbasin faucet' (lavabo bataryası), but the target translates it as 'küvet musluğu' (bathtub faucet), which is a different product." Suggested fix: lavabo bataryası.

3. Oil that became a nonsense word (English to Turkish, shipped)

• Source: Features such as the simplified liquid mercury introduction and automated oil purging make the Porion X the easiest-to-use mercury intrusion porosimeter
• Target: Basitleştirilmiş sıvı cıva doldurma ve otomatik yap boşaltma gibi özellikler, Porion X’in en kolay kullanımlı cıva intrüzyon porozimetresi olmasını sağlamaktadır
• Verifika: not flagged.
• What's Wrong: "The term 'oil purging' was mistranslated as 'yap boşaltma' (which is nonsensical) instead of 'yağ boşaltma' (oil purging)." Suggested fix: Basitleştirilmiş sıvı cıva doldurma ve otomatik yağ boşaltma gibi özellikler, ...

Our own test file#

The delivered files show what happens in the wild. For a check you can read in one glance, here is our own 14-segment English to Turkish file: 10 planted errors and 4 clean segments, one of them a trap. It is the file our end-to-end test runs on production. We ran three versions of it with different sentences and the same planted errors; every row below held in all three.

The table shows our result only: we have not run this file through Verifika.

Where Verifika is the better fit#

• The files must stay on your computer. The desktop app runs locally. If a contract forbids any online tool, that settles it.
• You want the tool to correct in place. Verifika fixes what it finds inside the file it opened. What's Wrong never touches your original: you accept the fixes you want and download a corrected copy.
• You want a quick check without an account. The web version offers one.
• Quotation marks and punctuation. Verifika is thorough here, and on one of our files it caught a quotation style we did not.

Where What's Wrong goes further#

• It reads for meaning. After the rule checks, a language model reads each segment next to its source and the segments around it, looking for mistranslations, omissions, additions and the wrong term or register. That is where the 22 errors came from.
• A second opinion on the rule findings that need a reading. A second model reads each critical or major rule finding that calls for a reading, such as a term, a number, a length or a checklist hit, and says whether it agrees, disputes it or cannot tell. The deterministic checks, such as a missing tag or placeholder, never go to it: same file, same finding. A disputed finding is never hidden; it stays in the report with the reason and counts at half weight until the reviewer decides.
• It tells you why. Each finding comes with a one-line reason and a suggested fix, so you can see when the model is wrong and say so. The decision is recorded.
• A score you can defend. The automatic score is 100 minus the penalty points per 1,000 words, with the arithmetic shown. The same defect repeated 40 times counts as one defect with a repetition weight, not as forty. On a small sample we show no score at all, rather than a number that means nothing. A reviewer signs the report off as pass or fail; the signature records their name, their role and the score at that moment.
• Spelling in every language. A dictionary proposes the candidates and the AI confirms each one before it is reported, so a rare but correct word form is not flagged. For a language no dictionary covers, the AI proofreads the segments itself.
• It works where the translators already are. Inside the editor in Phrase, XTM Cloud and Smartling, in Crowdin as an app, from a Dropbox folder, by email to a profile's own address, from an AI assistant through the MCP server or by uploading a bilingual file.

Which one to pick#

If your files cannot go online, or you want the QA tool to correct the file in place, Verifika is the better fit and this page will not change that. If what keeps you up at night is the sentence that reads well and says something else, run one file you have already delivered through What's Wrong and read the report next to Verifika's. The two reports will not look alike, and that is the point.

There is a free plan, no card needed. See the plans. If you want the longer version of how a model reads a segment, it is in What is AI translation QA.