bAIbelAV
/av/consistency11 / 11 · RESOLVED

Consistency

Two documents of the same job contain the same sentence. In a normal grid they are hundreds of rows apart in different files, so the two translations never meet, and the inconsistency is found by the client rather than by you. That is not carelessness. It is the grid.

So there is a second way to look at the same work: not in document order, but grouped by what repeats — whole segments, recurring phrases, and near-matches where the targets have drifted further apart than the sources did.

In numbers

183near-match pairs on two annual reports
13that were worth looking at
60–93%what the obvious version of the check would have flagged instead

Harmonisation

Every copy of the same sentence, side by side, across the whole delivery.

Harmonisation View abandons document order and groups segments by shared source, across every file of the job. Each set is marked as consistent, differing only in spacing or case, or differing; an untranslated copy beside a translated one counts as differing.

THREE FILES, IN DOCUMENT ORDER manual.docx Unplug the device before cleaning. Contact customer service. Keep the receipt as proof of purchase. warranty.docx Keep the receipt as proof of purchase. The warranty lasts two years. faq.docx Contact customer service. Unplug the device before cleaning. ONE SET PER REPEATED SENTENCE ORDER: MOST DIVERGENT FIRST Keep the receipt as proof of purchase. Translations differ manual.docx Conservez le ticket comme preuve d’achat. warranty.docx Gardez le reçu comme justificatif d’achat. Contact customer service. Differs in spacing/case manual.docx Contactez le service client. faq.docx Contactez le Service client. Unplug the device before cleaning. Consistent manual.docx Débranchez l’appareil avant de le nettoyer. faq.docx Débranchez l’appareil avant de le nettoyer. AN UNTRANSLATED COPY BESIDE A TRANSLATED ONE COUNTS AS A DIFFERENCE

Shared phrases, not just whole segments.

It works below segment level too: phrases of five words or more that recur inside different sentences, across the files, and are not already translated the same way everywhere. Pick one and the grid filters to its occurrences.

SHARED PHRASES · MIN LENGTH 5 · ACROSS THREE FILES the right to terminate this agreement 6 tokens · 4 segments · 3 files shared target wording: A 2 · 2 shares none supply-terms.docx The Supplier reserves the right to terminate this agreement without notice if the Customer fails to pay. Le Fournisseur se réserve le droit de résilier le présent contrat sans préavis si le Client ne paie pas. A service-levels.docx If availability falls below 99%, the Customer has the right to terminate this agreement by written notice. Si la disponibilité tombe sous 99 %, le Client a le droit de résilier le présent contrat par notification écrite. A renewal-letter.docx After the first year, either party may exercise the right to terminate this agreement on 30 days’ notice. Après la première année, chaque partie peut mettre fin au présent accord avec un préavis de 30 jours. – renewal-letter.docx Nothing in this letter affects the right to terminate this agreement under clause 14. Rien dans ce courrier n’affecte le droit de résilier le contrat prévu à l’article 14. – FOUND WHEREVER IT FALLS IN THE SENTENCE · WHICH WORDING TO KEEP IS YOUR CALL

Near-matches, without the noise.

Flagging two near-identical sources whose targets differ is useless, because the sources differ too. So the check asks whether the target differs by more than the source does: on two annual reports, 183 pairs diverged and 13 were worth looking at.

See it in the applicationHide the screenshot
That the question asked is comparative — does the target differ by more than the source — which is why the list is short enough to work through
Near-match harmonisation — a pair whose target differs by more than its source does, beside a pair that was correctly left alone

We built the easy version first, measured it, and threw it away.

The obvious version — "sources nearly identical, targets differ" — fires on 60% to 93% of every band. We measured it, found it would hand you a worklist that was about 90% correct work flagged as a problem, and threw it away.

Several documents, one grid, one index.

Open four files and you get one continuous grid, not a stack of tabs, each row still knowing which file it came from. A sentence that appears in file one and again in file three is marked in both.

Translation