An auditor does not need a promise. They need a record they can check. This page says what SazAI records, and where the record ends.

What gets recorded

Per step: what went in, what came out, and who asked

Every run records, per step, which file went in, what came out, what failed and why, what it cost, and which account asked. Lineage is a column in the database, so this file came from that one is checked by query. Failed work is recorded, and billed like work that passed.

1,901Comparison findings, counted across the four formats267 in PPTX, 167 in DOCX, 338 in XLSX, and 1,129 in PDF, in fourteen comparison runs. Not a sample: the query counted every finding in the database.
Measured against our own database on 2026-08-28. The store was rebuilt in October 2026; the count will be re-measured on the new store and dated here.

The full trace of one document is in your documents.

Where the record stops

At the language model, and we say so

Lineage follows each passage of text until it goes to the language model, and picks up again when it comes back. What happens inside the model is governed by the provider’s retention terms, not by our records. In an installation of your own, you choose the provider.

the provider begins here
covered by lineageupload, extraction, measurement, editing, outputoutside our boundarytext sent to the model follows the provider's terms
The line breaks on purpose.

What you can audit

Where the data lives, and what deletion reaches

The platform tables hold structure, measurements and lineage, not document content; the content lives in your organization’s own schema, and you check that by auditing the schema. Deleting a document is called withdrawal: backup copies can take up to seven days to catch up. The full account is in privacy and sovereignty; the decisions behind each capability, and what is still only direction, are in platform.

Where our engineering claims are weak

How far each one is anchored, stated by us

  • Our internal map of decisions and data reaches 55 of the 174 data tables; the repository reaches all 174.
  • Accessibility criteria (WCAG 2.2 AA) are cited in the design-system components, but no architecture decision records WCAG yet.
  • Testing has strong practice and guards that refuse at build time, but no external methodology declared.
  • Writing conventions rest on an internal, versioned contract, not an external standard.
  • Five references that shaped our thinking (W3C PROV, compound AI systems, GraphRAG, model routing, CommonMark) appear in no engineering artifact.
  • Market studies we cite support positioning, not engineering: executive surveys are perception, not audited financials, and preliminary or regional studies are named as such.

What not to use SazAI Corpus for today

Five limits, stated before you find them

  • Rebuilding structure. The engine changes content inside the existing structure; it does not create, delete, or reorder units.
  • Assisted creation as final delivery. It is the least mature operation and needs human review.
  • PDF translation into every language. Unsupported targets are refused or flagged at selection, never after delivery.
  • Automatic visual checks. The system counts elements; the last look is human.
  • The records as a certified audit. They are evidence of control; whether they meet a formal requirement in your sector is a case-by-case conversation.

When a limit goes away, it leaves this page with a date.

Prove it, do not take our word

Five things to demand before signing

The receiptof a delivery of yours, item by item
The lineageof which file produced what
The schemato audit where the data lives
The boundarydeclared, not discovered by you
The withdrawalexecuted, and what it does not reach
None of them asks for trust.

Upload a document of your own, translate and compare it, and get the count before and after. If one of the five cannot be done in your evaluation, that is our defect: tell us at contact.

What you see, ask, translate and create passes through here and leaves a trace.