The offer is a partition, not a list

What is offered, what is removed with a reason, and what nobody has measured yet

When you ask which languages the system translates, the answer that reaches you is not a vector of names. It is three sets, and the second is the one that matters for your decision: the targets offered to you, the targets removed with the reason for each removal, and silence, which means exactly one thing, and that thing is written down. This was a correction we made, and it is worth explaining why. For a while, a removed target simply vanished from the list, and an absence is the one answer that cannot carry a reason. Asked about a language nobody had measured, the agent answered that it *does not appear among the supported targets*. We never said it was not supported; we said nothing, and an absence has only one possible meaning to whoever reads it. Today the removal is a data point, with the reason attached.

offeredwith a caveat when there is one

removedand the reason comes with it

not evaluatednobody has looked, and that is not a no

A language outside the three sets above was not refused: it was not measured.

The difference is commercial before it is technical. Telling you we do not do your language when the truth is that nobody looked is a lost deal paid for with a missing field, and the loss falls on both sides.

Two different questions, and silence means the opposite in each one

There is a font that draws the alphabet, and there is knowing how to assemble the line

Your target goes through two independent questions, and confusing the two is the classic error of this domain. The first is whether a font exists that can draw those letters. The second is whether we know how to assemble the line: right-to-left scripts need to be reordered, and several scripts from South and Southeast Asia choose the shape of each letter by context. The two fail in opposite ways, which is why silence means the reverse in each. When a font is missing, the failure is loud: the segment is not replaced, the original sentence **stays on your page** in the source language, the result goes out marked as partial, and your receipt names what was left behind. You see that something did not happen even without speaking either language. So a target nobody has measured yet can still be offered to you: if it goes wrong, it goes wrong in plain sight. When assembly is missing, the failure is silent: every letter is drawn, your page looks finished, nothing is dropped, there is no receipt line to write, and only someone who reads that script notices that the line runs backward. You could take that file into a meeting with no way of ever knowing it was wrong. So on this axis, unmeasured is treated as unsupported, and the target is removed before it reaches you.

the fontfails in plain sight, and the receipt counts it

two questionsand both need to pass

assemblyfails in silence, so it is refused

Only one of the two failures can be reported by whoever suffers it. That is the one that decides the default of silence.

Where this bites, and it is only in one format

In office formats, the program that opens the file does the drawing; in PDF, we do

In the three office formats, the file stores letters, and the program that opens it assembles the line. That is why Arabic works in a presentation, a document, and a spreadsheet: the drawing is not ours. In PDF, the text is already drawn at a position; translating means redrawing, and there, the drawing is ours. The consequence is declared when you choose the target, not when you receive the delivery. Arabic is not offered as a PDF target, with the reason stated in the offer itself. Chinese, Japanese, Korean, Hindi, and Thai are offered in PDF with a warning. And your scanned page, which is an image and not text, comes through flagged instead of being chewed up in silence. One thing worth separating, because the two look the same and are not: the configured list of targets is an operating decision, and what we described above is a property of rendering. A language can be missing for either reason, and the offer states which one it is.

A warning without a size is not a warning

The same caveat has meant both "a few passages" and "the whole document"

The rendering caveat named which scripts were left out and never said how much of your document that was. Read by you at the moment of choosing, it sounded like *a few passages*, and in a real case it sounded exactly the same when the truth was *practically everything*: a PDF translated to Japanese came back with most elements still in plain English, wearing the same warning a document almost entirely in Latin script would wear. The fix is structural, not editorial. When your target language **is** that script, everything the job translates comes out in it, and the honest extent at the moment you choose is *all text translated for this target*. This became a field in the response you receive, rather than a sentence someone has to remember to pass on to you.

And today's state, dated. Translation has been down since 2026-08-23 on provider credit, and the last successful run was on 2026-08-20 at 22:43 UTC. This page describes the capability and does not invite a live test until that returns.

Limits per format are stated in [your documents](documents), and what we recommend against doing is in [trust](trust). What translation preserves when the language changes is in [preserved layout](layout).