Open a Turkmen file in memoQ with the default QA profile running and you'll drown before you translate a single confirmed segment. Hundred of warnings. Missing numbers that aren't missing. Terminology mismatches that are just Turkmen doing what Turkmen does — bolting six suffixes onto a root so the term base can't recognize its own headword. The checker isn't reading the language. It's pattern-matching against rules somebody wrote with English and German in mind, and it's telling your PM the file has 240 quality issues.
She reads that number. She doesn't read what's under it. I've written about that before. Today I want to be more useful and actually tell you which switches I flip.
The flags that lie for Turkmen
Terminology consistency is the loudest liar. Your term base has şertnama (contract). In the text it shows up as şertnamanyň, şertnama boýunça, şertnamalar, şertnamada. The checker sees a term that's "present in source but absent in target" and fires. It's there every time — it's just inflected, because that's how an agglutinative language works. memoQ's fuzzy term matching helps a little, but it still can't cover the full case-and-possessive spread. So the terminology check, the one feature agencies buy memoQ for in regulated work, becomes a source of false alarms exactly where you'd most want it clean.
Length checks are the second problem. Turkmen runs long against English. A UI string that fit in 42 characters in source blows past its limit in target not because I wrote something wrong but because the plural and the locative case each cost me letters. The length warning is real information for subtitling and UI — I don't kill it — but I read it as "go look," not as "you made an error."
Number-format checks fire on the decimal comma versus point, on non-breaking spaces in figures, on date order. Half legitimate, half noise depending on whether the client localized their number conventions or froze them. And the double-space and trailing-space checks catch things that are genuinely mine to fix, so those stay on. Cheap, honest, never wrong.
What I actually keep on
Tag verification, always. If source has an inline tag and target doesn't, that's a broken build downstream, full stop, no argument with the language. Turkmen doesn't get a vote on whether a <b> closes. Same for placeholders — %s, {0}, {{count}}. Drop one of those and you've shipped a bug, not a translation. This is the check that earns its keep. It doesn't care what language you speak; it cares whether the string will render.
Inconsistency in target — same source segment translated two different ways — I keep on and I actually welcome. In a 12,000-word oil-and-gas manual, that flag is the one that catches me having called a valve two different things on page 3 and page 40. That's a real find. The checker's dumb, but on this narrow thing it's dumb in my favor.
Everything terminology-related I turn down to a warning I can batch-review at the end, not a real-time interrupt. And I tell the PM up front that the terminology error count in the report is inflated by inflection and shouldn't be read as a defect rate. Sometimes she listens. Sometimes the number goes to the end client anyway and I get an email.
The check that isn't a check: live preview
Here's my actual opinion. The most valuable QA feature in a modern CAT tool isn't the rule engine at all. It's the live preview. memoQ shows the segment in final layout while you translate. Smartling and Crowdin do it in the browser. Trados has caught up. That in-context view does more real quality work for Turkmen than the entire automated profile, because most of my genuine errors aren't rule-shaped.
A rule can't tell you a sentence reads stiff. It can't tell you the case ending you chose is grammatically fine but makes the phrase mean something slightly off in this button. It can't see that your long string wraps and eats the icon. Preview shows you the string sitting where a human will read it, and suddenly you catch the thing no check would ever flag. The industry keeps selling adaptive MT that "learns from human feedback" as the QA story — Smartcat's agents, Phrase's rewriting, Trados's self-learning NMT. Fine. But for a language with almost no clean feedback to learn from, one careless post-editor's corrections train the model wrong, and now the machine is confidently reproducing somebody's mistake. The feedback loop only raises quality if the feedback was any good. For Turkmen, that's a big if.
So I trust my eyes in preview over a score, and I trust tag verification over both. The rest I tune down until it stops crying wolf. Because a QA profile that flags 240 issues on a clean file has taught everyone who reads it to ignore the profile — which means the day there really are 240 issues, nobody looks.
Give me a checker that shuts up unless it's right, and I'll actually read it.