2026-09-11
The analysis of the package of files before translation now also writes down ready wordings for everything that repeats to a pattern: "Figure A.1", "Table B.2", "Version 3", document and product codes. During translation those wordings are taken from that list. Until now such a decision was made again in different batches, and the name of one and the same figure could come out in two different wordings in neighbouring files.
How we tested it. We took three files of engineering documentation where the drift was clearest: 2,347 paragraphs, about 11,700 words. We translated them twice, the old way and with the wordings written down. Of 40 groups of numbered names, 12 came out inconsistent in the first translation and none in the second. Product codes that have to stay exactly as they are in the source were changed 39 times in the first translation and never in the second.
Consistent is not the same as correct. If a wording is a poor choice, it is now consistently poor across the whole set, so your glossary still takes precedence: where a term is in it, the translation comes from there. We also checked whether this crushes legitimate variation. On that set we found one case: two names differing only in singular and plural merged into a single English wording, which for English is correct.
The analysis before translation now takes longer, and it runs once per task. Prices are unchanged.
2026-09-09
Files are deleted from our servers automatically after 7 days. That deadline is now visible on every task: in your account each one carries the line "Files will be deleted in N days", and once the period is over it reads "Files deleted", with no download button next to it. Open an old result link and the service tells you plainly that the files for that task have been deleted.
What happens to the text while the task is alive. Our service logs contain no document text: they keep events and errors, not the contents of your files. Your text is never used to train models — neither by us nor by Anthropic, the model provider. Anthropic retains API data for a limited period under its policy and does not train on it. We do not control retention on the provider's side.
To have documents deleted earlier, write to [email protected] and give the task number or the account address. We delete within one working day and confirm by email. There is also a point worth reading before you upload: if your agreement with your client forbids passing the document to third-party data processors, do not upload that document to the service, ours or any other cloud one.
None of this changes prices or translation times.
2026-09-04
If you upload several files in one task, identical paragraphs in them are now translated once for the whole batch. Until now, repeats were collapsed within a single file only: a paragraph that appears in ten documents of a package was translated again in each of them. Now repeats across files are translated once and the translation is placed in every file of the task, including across files in different formats — DOCX and XLIFF, for example.
What it adds up to. In our statistics on multi-file tasks, repeats inside files come to 26% of the words; counted across the whole task, 44%. The difference between the two is the cross-file repeats: 17% of all the text was being translated again only because the same paragraph sat in a neighbouring file. How much of it there is depends on the package: little in a set of unrelated files, a lot in uniform sets (a series of equipment data sheets, a set of contracts).
A repeat has to match exactly. If two paragraphs differ by a single character, they are different segments, and each is translated separately. Collapsing does not work on PDF, where the same text sits in different layouts, and the discount does not apply to PDF either. The memory of repeats lives inside one task — the same files uploaded tomorrow as a separate task will be translated from scratch.
Repeated segments carry a 75% discount, now including the ones that repeat across files. The discounted amount is shown in the estimate before payment, with repeats on a line of their own.
2026-09-02
Since HardTrans was launched, every document has been analysed as a whole. Now we have added a statistical check alongside the LLM check. Before the analysis, the system builds a frequency list of terms: which word occurs how many times across the whole batch of files. That list goes to the LLM at the analysis stage, and the model reads it together with the files. The decision on each term is still made once, before translation starts, but now it is pinned more firmly, so the model is far less likely to reach different decisions in different batches.
How we tested it. We took an 800-paragraph slice of an engineering specification and translated it two ways: as before and with the inventory. Quality was scored on our review form: 100 points minus a penalty for every error found. The score went from 82.8 to 87.5. To put that in perspective: two translations of the same file made the same way differ by about 0.6 points, which is the model's natural spread. A gain of 4.7 points is seven times that spread, so it cannot be chance. We also tracked one term separately: it used to be rendered by two different words and drew 47 findings; now it is one word and 17 findings.
Your glossary still takes precedence. If you upload one, the translation comes from the glossary, and the inventory covers only the terms the glossary does not have. Nor does it make a translation error-free: it settles which word is used throughout, not whether that word is the right one.
Translation takes slightly longer: the analysis with the inventory needs more time, about a minute and a half on an 800-paragraph document. Prices are unchanged.
HardTrans