Benchmark sheet
Technical manuals, translated and typeset. This is the tool's sheet, published exactly as it looks right now. The product page is here.
⛔ One of the seven rows failed, and it is the one that matters most
A page that holds only a table is dropped. The tool keeps the pages it detects as Romanian, and a specification table — part numbers, volts, amps, kilograms — contains no Romanian words to detect. Its language score is 0.0000 against a threshold of 0.012. Not near the line. Nowhere near it.
We then put the table between pages of prose, the way it sits in a real manual: pages 3, 4 and 5 of 8 were dropped, and neither „Tensiune" nor the part code „A-101" reached the output. In a product manual, that table is often the page the customer actually needs.
It does not crash and it does not warn: the caller can see which pages were kept, but the text simply comes back without them. That is the worst shape a fault can take, and it is why the row above is red on our own product.
What is measured
Standard benchmark — the documente-v1 set, seven files
Each row is filled in when we push the file through the tool, on our own machine, and write the date. Nothing is marked „passed" from memory or from habit.
The row that decides whether the sheet gets published
What we are doing about it
- The red row stays red until the tool changes. Not until we word it better. A sheet you can only pass is not a sheet.
- The fix is known and small. A page with no detectable language that sits between pages of Romanian belongs to the Romanian block. Tables do not have a language; the pages around them do.
- We found this on our own product, with our own benchmark. That is the whole point of the set being fixed and published. Download it and run it against us.
- Your sheet will look the same, if you sell here. Same shape, same rows, same rules — including the red ones. How goods reach the shelf.