AI supply chain
Model Integrity
Available now
Sign, verify, and trace every model artifact that reaches production — using the same signing pipeline that protects our own releases.
Shipping today. Ask for access and we will get you a build.
What it does
Signed model artifacts
ML-DSA signatures over weights, adapters, tokenizer, and config. A file altered anywhere between training and serving fails verification instead of loading quietly.
Provenance you can replay
A manifest recording which base model, which dataset revision, which pipeline run, and which approver produced the artifact you are about to serve.
Verification at load time
A gate in the serving path that refuses unsigned or mismatched artifacts, so the check happens where it matters rather than in a policy document.
Registry-agnostic
Works against a public hub, an object store, or an internal registry. Nothing about the signing scheme assumes where the file lives.
What gets signed
Everything the serving path loads, not just the weights file.
- Weights & adapters
- Per-file signature
- Tokenizer & config
- Per-file signature
- Training manifest
- Provenance record
- Container image
- Digest pinned
- Signature algorithm
- ML-DSA-65 (FIPS 204)
How it goes in
Each step is a checkpoint you sign off before the next begins.
- 01
Register
Enrol the model and the pipeline that produces it against a signing key held in your own custody, not ours.
- 02
Sign
Your CI signs the artifact set at the end of a training or fine-tuning run and publishes the manifest alongside it.
- 03
Verify
The serving path checks signatures before load. Anything unsigned, altered, or unexpected is refused and reported.
- 04
Audit
Every load is recorded against its manifest, so you can answer what was serving on a given day months later.
Works alongside
Nothing here needs the rest of the line to be useful — but they are built to fit together.