Posting a working reference example I built against the NIM-hosted Nemotron API: an agent’s proposed action gets evaluated through a policy check before being treated as authoritative, then the decision is signed with ML-DSA-65 (FIPS 204), independently verifiable by anyone with the public key, not just trusted on the agent’s own report.
I want to be specific about scope, because I think this matters to anyone else building here. NemoClaw and OpenShell already give agents policy-based guardrails and sandboxed access so behavior stays predictable and auditable. This isn’t a competing runtime. It’s one layer beneath that: a cryptographic signature on the attestation record itself, so it stays independently verifiable even after today’s classical signatures become forgeable, which matters for anyone whose agents will still be running in production a decade from now.
What’s actually in the repo:
- Real ML-DSA-65 signing, verified round-trip (not just “names the right algorithm”)
- A policy check that denies malformed/low-quality actions outright and marks everything else
require_approval, never auto-executed - Runs entirely on CPU, the model call goes to NVIDIA’s free hosted NIM endpoint, no local GPU needed to try it
Repo: GRC_Claw/examples/6g-nemotron-attestation at main · AAH20/GRC_Claw · GitHub
Genuinely interested in feedback from anyone on the AI-RAN or Aerial side on where this would actually need to plug into NemoClaw’s guardrail flow to be useful in a real deployment, that’s the part I don’t have visibility into from outside NVIDIA, and it’s the thing I’d want to get right before extending this further.
Background briefly: I’ve spent this year finding and fixing a non-functional post-quantum implementation in my own open source stack (documented publicly, not something I’m glossing over) and building the governance/attestation layer now live on my own platform. If anyone here is working on the trust and attestation side of AI-RAN specifically and wants to go deeper on this, I’d welcome that conversation directly.