Quality, languages and privacy

One product

There is one quality level, and it is called Verified. There is no fast lane, no extended lane, and nothing to upgrade to. Every order gets the same nine files and the full quality-control report.

There used to be three tiers, and all of them shipped exactly the same files. A ladder that changes nothing sells nothing — so it was retired, at the middle tier's price anchor. We did not raise the price.

What gets checked

• Terminology consistency — that the same term is rendered the same way from the first page to the last.
• Structural completeness — that headings, tables, lists and captions did not vanish or merge.
• Numbers, formulas and citations — that they did not shift and were not translated where they must not be.
• Reading order — that the document reads in the order it was written, even where the source was two-column or scanned.

What was checked and what was flagged reaches you as three files: the quality-control report, the translation protocol and the terminology sheet. The check is not an assertion — it is a document.

Translation for reading and translation for submission

To read a paper, understand it and work with it — the translation is ready as it is. To submit translated text under your own name, publish it, or rely on it legally — that needs your own pass over it, and in some language pairs a domain expert too. The terminology sheet and the bilingual view exist precisely for that pass.

Four coverage levels — not every language pair is equal

Designed for 50+ languages; verified pairs first (English⇄Hebrew). Every language pair carries a coverage level, and it is set by what has actually been tested — not by how widespread the language is.

Controlled

The pair passes a fixed control set, and every change to the system is checked against it before it ships. English⇄Hebrew.

Controlled with uncertainty

A best-effort translation for this pair, without a control set yet. Uncertainty is flagged explicitly in the quality-control report and in the terminology sheet, rather than being absorbed.

Reading only

A best-effort translation routed through English as an intermediate language. Intended for reading and understanding — not for submission and not for publication.

Experimental

We can attempt this pair, and we make no quality promise. That is said in advance rather than discovered afterwards.

Verified is the name of the product: it describes what is done to every order. The coverage level describes the language pair. A Verified order in an experimental pair is still an experimental pair — it gets the same nine files and the same checks, and that does not turn it into a controlled pair.

Your documents are not raw material

Your document is sent in order to produce your order, and for nothing else. What follows is not a statement of intent — it is what the system does, line by line.

• Your file is uploaded straight from your browser to storage, and downloaded the same way. It does not pass through our web servers.
• A scanned page is first read by text-recognition software running on our own servers. At that stage the page does not leave our processing environment.
• The document's layout parsing also runs in our environment. No third party sees the document in order to parse it.
• Before you pay, only a bounded excerpt is sent to a model — at most 8,000 characters. Not the whole file.
• After you pay, the full text is sent, in batches, to produce the translation.
• Deliverables are delivered as time-limited download links, not as email attachments.
• The pre-payment analysis artefacts sit in a private bucket customers cannot download from.

• Auto-deleted after 14 days — or delete anytime yourself
• Deletion is automatic and runs daily.
• It removes the file you uploaded, the deliverables, and the intermediate analysis of the document as well.
• You get a reminder email before the deletion, and you can delete everything yourself at any time from your download page.

Your document is processed by models from OpenAI, Anthropic and Google. This is not a footnote — it is part of how the product works, and you should know it before you upload.

Who receives what

• OpenAI · Anthropic · Google — the document text, and with Google also rendered page images.
• Supabase — the storage and the order database.
• Modal — the compute environment everything runs in.
• Vercel — the website and the order flow. Metadata only; not document content.
• Resend — your email address and the download links. Never the content.

Under our AI providers' API terms, your content is not used to train AI models.

Not used for training is not the same as not retained. The providers keep logs for a limited period for abuse prevention, and content flagged by trust-and-safety systems is retained by them for longer.

Files are encrypted in transit and at rest. This is a property of the storage and compute platforms we use.

Storage and the database are located in Japan.

Medical and legal documents

We translate medical and legal documents, and we hold them to a stricter accuracy standard than other material. That is a translation-quality measure, not a privacy measure. A medical document travels exactly the same path as any other: it is processed by the same third-party AI providers, and no identifying information is removed from it before it is sent. If sending a document to external AI providers is not acceptable for the material you hold, please do not upload it.

Where translation-memory reuse is enabled, it is strictly limited to your own account and keyed by a hash, never by your email address. It is off by default.

We do not claim to have signed data-processing agreements with our providers, because we have not. What we have is the list above — who receives what, and what their terms say.

The full privacy policy

• A number you can point at. Every number on this site — nine files, 14 days, $3.90 — is what the system actually does. A number that cannot be pointed at does not appear.
• The limit is stated in the same size as the claim. There is no fine print here, because the limits are part of what we are selling.
• What has not been tested is said not to have been tested. The four-level model exists so that a pair which has not passed control does not pretend to be one that has.

If something in a deliverable is wrong

Our commitment: free technical reprocessing if a technical problem affected the deliverable, within 14 days of delivery.

What qualifies

• Text that was omitted, left untranslated, or placed in the wrong location because of a processing fault.
• A deliverable that failed to generate, or is corrupt.
• A deliverable that does not match the order — wrong target language, wrong scope.

What does not qualify

• style preference
• human proofreading
• certified translation
• handwriting fixes
• perfect reconstruction of low-quality scans
• manual chart editing
• text inside figures, charts, axis labels and legends — it is never translated, and that is disclosed behaviour, not a fault.
• layout fidelity and the searchable PDF are best-effort, not a guaranteed reproduction.

How to ask

Through the contact form, with the subject "Problem with a deliverable / reprocess request" and your order number. Within 14 days of delivery — after that the source file has already been deleted, and a re-run is impossible.

We reprocess first. If the re-run cannot fix it, or the order could never have been delivered as described — you are refunded, in full. There are no partial refunds: the product is sold as one package at one price.

The full refund policy

Contact

Everything on this page applies from the first document.