Your documents, opened: what the engine read, what you edited, and the receipt of every change.

The unit of your format

The conversation changes word because the unit changes nature

A presentation has slides. A document has sections. A spreadsheet has sheets. A PDF has pages, which are already the final rendering and not a space still to be laid out. The screens and the agent use the word your format uses, not the word of the most common format.

Text documents are the one case where two words are needed, and it is worth saying why. A document is counted in pages, which is what you see and what you would say out loud. It is analyzed by section, and a section here is a layout break: page size, orientation, margins, headers, columns. A document with no explicit break has a single section, so for an ordinary company document the deep exam runs per document, not per page. That difference is stated, not smoothed over.

The X-ray, whole

One reading, on entry, and everything the package carries gets an address

An Office file is a package, and SazAI Corpus reads the whole package, once, through an engine that reads the format specification and does not interpret. The X-ray is what comes out of that reading, and it is more complete than the presentation showed.

Each piece of text stays tied to the slide and the shape it came from, with the formatting it carries. Comments come with author, anchor, reply thread, and the two states that matter: resolved and with a task. The speaker notes come in full. Links, images, and hidden slides all appear, and the slide someone hid appears for you too, marked as hidden. The properties say who created it and when it was revised.

The sentence someone pasted into the seventh presentation, and nobody knows where it came from: the X-ray knows. That is what makes it useful for audit, because every statement in the document has an address, and the address belongs to the file, not to our opinion about it. Nothing beyond what is in there.

What the engine measured

A deck knows it is repeating itself before you do

The engine measures the whole document and stores the result under the lineage of that run, that is, which file produced what. The signals have names:

Deterministic, and with no opinion. The engine tells you the third and the ninth presentations are nearly the same; what to do about it is a conversation with the agent. The measurements for each document stay available for reading across documents, so a whole portfolio can be read without opening one file at a time.

Editions and comparison

Three relations declared before the first number

Your proposal is on its fourth version and nobody remembers what the third one said. Every edition of yours, translation or revision, enters the vault as a complete document, with its own X-ray and measurement. Compare puts your original in the first column, always, and declares the relation between the two before showing the first difference: translation, revision, or independent. With no declared relation, the comparison does not appear.

And what the comparison gives you back is not an impression: in a 45-slide edition, 633 elements before and 633 after. The text changed; the structure did not. Counted, not estimated. The difference between this and opening two files side by side is that the result stays on record per unit and can be reopened later, by someone else, without redoing the work and without depending on whoever ran the comparison.

Original and derivative stay linked

The relation survives the departure of whoever created it

When one document produces another, the link is recorded, and the derivative is a complete node with the same treatment as the original: it is read by the engine, measured, and navigable on its own.

This solves a problem that looks small and is not. In a folder, the original and its five translations are six files that only one person knows how to relate, and only until they change teams. Here they are one original with five declared derivatives, and nobody needs to remember anything to reconstruct the relation.

The record is not written afterward by someone who observed the operation. It is born from the operation itself, which is why it cannot disagree with it. Every one of these links is also a synapse in your vault, and the whole shape they draw is the network your vault becomes.

What the receipt proves, and what it does not prove

Evidence of control, stated with its boundary alongside

The receipt proves that the operation happened, on which version, with which parameters, and what it changed. A translation records completeness per segment, the glossary terms applied, and what did or did not make it into the final file. A rewrite records which fields were changed and which were kept.

It does not prove the result looks good visually, because the system measures structure and does not look at the final page. It is also not certified audit: whether this history meets a specific audit requirement in your sector is a question we answer case by case, not a claim this page makes.

A skeptical buyer usually puts the question directly: the AI generated something, and now, how do I know I can use it. The answer we prefer is not to ask for trust in the model. It is to show what happened to the file.