Any tool returns a file in another language. Here, three things come back with it, and none of them is optional.
The receipt travels with the delivery
What changed and what stayed, item by item, kept and reopenable
Every translation arrives with its own comparison: what changed and what stayed, unit by unit. You do not need to believe nothing broke: you see it, and so does whoever opens the file after you. It is the same receipt that every delivery carries, applied to translation.
| Slide 12 · title | translated | ✓ en → pt-BR |
| Slide 12 · chart | intact | = unchanged |
| Slide 12 · speaker notes | translated | ✓ en → pt-BR |
| Slide 13 · image | intact | = unchanged |
| Layout · fonts · formatting | preserved | = unchanged |
The translated version is a document, not an attachment
It enters the vault next to the original, with its own X-ray and measurement
The translation enters your vault next to the original as a complete document: read, measured, navigable on its own, and linked to the original by a lineage that does not depend on anyone’s memory.
When someone asks where the version sent to the client came from, the answer is recorded, and comparing the two is one click. More in your documents.
Your glossary
One spreadsheet goes up and holds for the whole delivery, in every language
Upload your glossary, a simple spreadsheet, and product names, legal terms and the jargon of your field come out the same from the first slide to the last, in every language. It stays versioned in your vault and applies to every job that requests it.
And the boundary is checked, not presumed
The model's text enters counted, and a difference fails the step
When translation puts a language model’s text into your file, every segment goes out delimited by markers, and what comes back is counted against what went out. If the model drops or invents a segment, the step fails instead of breaking the document, and you find out before the meeting. The count checks completeness, not the quality of the translation. How the count works is in the conversation is with the AI.
And today's state, dated. Translation stopped on 2026-08-23 when the provider credit ran out; the last complete translation job finished that day at 02:26 UTC. The credit was restored on 2026-09-30, and the next complete run will date this line. This page describes the capability and does not invite a live test until that run exists.
Writing and script limits per language are in languages; what stays in place when the text changes is in preserved layout.
When an AI reads your files
Yes, AI reads. The question is whether you can prove what it read.
| SazAI Corpus |
|---|
| records what it read from each file, and lists the parts and content it detected and did not read |
| gives the place in the document, and the file’s hash when one was recorded |
| reports an absence from a search, with what the search covered |
| quotes the norm from an official source with the wording time and source details the result provides: word for word where the licence allows, and as a marked summary where it does not |
| gives the evidence and asks the AI to leave the conclusion to you |
| will be installed on each client’s own server: the library will be stored there, and the text an AI reads goes to that AI’s provider |
Each client’s SazAI Corpus will be installed on that client’s own server, and the library will be stored there; this site is our showcase. The text an AI reads, whether it translates or answers, goes to that AI’s provider. The assistant you ask is the one you chose, under your own account.