The worst error I ever shipped was a 100% match I never opened. Oil and gas, a safety procedure, and the TM served back a segment from an earlier job where someone had swapped "shall not" for "shall" in Turkmen. One word. Locked, green, zero penalty on the grid. Studio told me it was perfect. The client's HSE reviewer disagreed, loudly.
So let's talk about the match percentage, because the number is one of the most quietly dishonest things in our trade.
What the percentage actually measures
When Trados or memoQ tells you a segment is a 100% match, it means the source text is identical to something already in the memory. That's all. It says nothing about whether that prior translation was any good, whether it was approved, whether it fits this document, or whether the previous translator was fighting a deadline at 2 a.m. and guessed.
A context match — CM in Trados, 101% or 102% in some setups, ICE in the old Wordfast world — goes one step further: the segment before and after also match. That's genuinely stronger. But stronger than "identical" is not the same as "correct." I've seen CM segments carry a terminology choice the current client explicitly bans. The context was identical. The instructions weren't.
Here's the part that makes me grind my teeth. On the discount grid, 100% matches get paid at 0–25%, and plenty of agencies lock them outright so you physically can't edit them. Context matches, same story. The logic is: this is already done, don't touch it, don't bill for it. Fine. Except the delivery still goes out under my QA, and if the locked segment is wrong, nobody's going to say "well, that one was locked." It's a Turkmen file. I'm the Turkmen. It's mine.
The segment you don't look at is the one that burns you
I've made a rule over the years: I read the 100% matches. All of them, at least once, at reading speed. Not to re-translate — to catch the ones that don't belong.
The usual suspects:
- Numbers and units riding inside a matched sentence. The source changed from 50 to 500, the fuzzy logic still called it a 100% because the digits are tokenized as placeables, and the TM happily reused the old figure. This is why I never kill the number-verification QA check even when it screams false positives all day.
- Gender, case, and agreement in Turkmen that depended on the surrounding sentence. A match that was correct in its original paragraph can land grammatically wrong three documents later. The database doesn't track that. My ear does.
- Client-specific terminology that drifted. Company A calls it one thing, Company B another, and the TM is shared across both because some PM merged two projects to save space. The match is 100%. The term is wrong for this client.
- Formality and register. A UI string translated informally two years ago comes back into a legal-ish context. Identical source, wrong voice.
None of these trip the automatic QA. The QA panel checks tags, numbers, consistency, forbidden terms, double spaces. It does not read for meaning, and it never will, no matter how many LLM-judge features get bolted on. A number check catches a changed digit. It does not catch a correct digit attached to the wrong noun.
How I handle it without eating the cost
I don't pretend to re-edit locked content for free. But I do two concrete things.
First, before I accept a job with a big leveraged chunk, I ask the PM one question: is the TM approved and final, or is it "whatever was in there"? The honest PMs tell me straight. If it's a legacy memory of mixed origin, I say I'll flag issues in the 100% band but I'm not warranting them unless I'm paid to review. Put it in writing. That one line has saved me from arguments I'd otherwise lose.
Second, I keep a running list of the matches I distrusted and corrected anyway, with comments in the file. If a locked segment is clearly dangerous — a safety instruction, a negation, a dosage, a legal "shall" — I unlock it, fix it, and comment why. Then I tell the PM I overrode the lock on segments X, Y, Z for these reasons. Nobody has ever been angry about that. They're angry about the opposite.
The grid treats a 100% match as finished work. The QA tool treats it as clean because the source hasn't changed. Both are looking at the database. The reader is looking at the sentence. Those two things agree most of the time, which is exactly why the gap is so easy to miss.
Treat every leveraged match as a claim someone else made, not a fact. Read it like you'd read a stranger's work — because that's what it is, even when the stranger was you, eighteen months and a thousand segments ago.