Training data that proves its own composition.
Regulators and customers increasingly ask the same question of an AI dataset: what is it actually made of, and can you prove it? AT-1 Corpus Governance seals a dataset's real-vs-synthetic composition and per-record lineage into one verifiable artifact, so a value tamper or a composition lie is detected, not taken on trust.
Per-record lineage
Every record carries where it came from: real source vs synthetic generator: so the corpus can answer 'what is this made of?' down to the row.
Composition you can prove
The exact % real vs synthetic is sealed into the artifact. A reviewer verifies the mix without trusting your word for it.
Tamper- AND lie-detected
Alter a value, or misreport the real/synthetic split, and verification fails. It catches both data tampering and composition misrepresentation.
EU-AI-Act ready
The provenance record high-risk AI obligations ask for: kept as one verifiable, exportable artifact, not a spreadsheet of promises.
Where we're honest
This isn't a claim that synthetic data is as good as real, it's a claim that you can prove what your corpus contains. Validated on real public datasets; it catches both tampering and a misreported real/synthetic split. The engine ships compiled and license-gated and runs on your own machine, your data never leaves it.
The question an auditor actually asks
Not how large the training set was. What was in it, where each part came from, how much of it was generated rather than observed, and whether any of the evaluation data leaked into the training data. Those are four separate questions and a spreadsheet answers none of them in a way anyone can check.
What the corpus carries
- Per-record lineage. Every record is tagged with where it came from and whether it is real or synthetic, so a composition claim is computed from the corpus rather than asserted about it.
- A sealed content root. The composition is signed and bound by hash to the data it describes. Change a record and the manifest stops matching.
- A contamination check. Run your evaluation set against the training corpus and get the overlap. If your benchmark numbers are good because the answers were in the training data, this is the tool that tells you before a reviewer does.
Honest scope, and it matters
This proves the corpus you assembled and declared. It does not prove what a training run actually consumed. Nothing outside your training pipeline can prove that, and anyone claiming otherwise is selling you a document rather than evidence. What it removes is the gap between what you say your corpus contained and what it provably contained, which is the part a regulator can check and the part you can be caught on.