AI TRAINING DATA
Provenance
Proof of what trained the model.
The problem
One unlicensed training batch can expose a model developer to legal and reputational risk. Most teams can't prove, after the fact, what went into a checkpoint.
What it does
It matches each batch to its declared source and license before ingestion, then anchors a training receipt tied to the checkpoint that consumed it.
What it deliberately does not do
It never grants a license or clears rights. It never asserts that a model's output is free of infringement.
What touches the chain
Onchain: a fingerprint of the record, plus minimal provenance (when, by which wallet, under which method version).
Off-chain: the underlying data. It is never published.
Correction model: a fix is added as a new version, so the history cannot be quietly overwritten.
Who controls it: receipts live in a registry controlled by your own wallet. NVNM is not a gatekeeper.
How it fits your workflow
Source and curate batches · Check each batch against its license · Anchor a receipt at training time · Keep a record of what trained each checkpoint.