I Built an Operating System Because I Don't Trust AI Agents to Follow the Rules
Third research paper published this year. This one goes all the way down to the operating system.
I've spent the last year developing a research program around a problem I believe will become increasingly important as autonomous AI systems move into enterprise production: Who actually controls what an AI agent is allowed to do—and how do we prove it?
My first paper, Agentropy, explored the theoretical foundations of autonomous agency.
My second, Git Ontic Trace (GOT), addressed verifiable computational causality and execution provenance.
Today I published the third: Odin OS, an experimental bare-metal RISC-V operating system that explores enforcing cryptographically authorized memory access at the processor boundary.
The initial prototype is implemented in Rust, with adversarial tests, formal verification artifacts, and reproducible evidence.
The commercial thesis behind this research is straightforward:
As AI agents gain the ability to execute transactions, modify infrastructure, and operate autonomously, enterprises will need more than guardrails, prompts, and audit logs.
They'll need infrastructure that can enforce authority and independently verify execution.
I'm building toward that infrastructure through Ravenhelm.
The research is public because I want the underlying claims tested, challenged, and improved.
I'm particularly interested in connecting with founders, investors, and technical leaders working on agent infrastructure, enterprise AI security, and verifiable autonomous systems.
Odin OS paper and research artifacts:






Comments4
Making the Odin OS research public so others can test and challenge the claims is a smart way to earn trust from cautious enterprise buyers. A company letting an AI agent move money or change its systems wants proof that the limits hold even when the agent doesn't follow the rules, and that is exactly the worry you're going after.
Ultimately I'd love the opportunity to have my work truely peer reviewed and published more broadly.
For the peer review you are seeking, a claim-to-evidence map would help a reader choose a bounded starting point: the exact property, assumptions and trusted components, then the artifact or test supporting it and what remains outside scope.
Which single claim would you most like an independent reviewer to try to falsify first? In particular, where do you draw the boundary between enforcing an authorized memory access and establishing that the resulting business action is appropriate? Those seem like distinct review questions.
I have read this post and the exchange, not audited the paper or prototype. I would keep that distinction explicit in any feedback.
@Rhode Brunel Moungondo Thank you for this - genuinely. I've never had anything peer reviewed before, and this is only my third publication on Zenodo, so I'm learning as I go. Your comment was the single most helpful feedback I've gotten so far. You asked the exact questions we needed to answer before this could ever stand up to real scrutiny. Because your critique deserved more than a short comment, I put together a dedicated follow-up post walking through both answers (narrowing our primary falsification target to ODIN-C001 unauthorized-byte preservation, and drawing the PEP vs PDP boundary at ODIN-C012): ai.lounge.ai/p/from-prototype-to-peer-review-answering-rhode-brunel-nov-9-2026 We also published the 14-claim registry, 12 falsification protocols, and opened the repo at github.com/nwalker85/odinos. Grateful for the direction.
4 comments
Making the Odin OS research public so others can test and challenge the claims is a smart way to earn trust from cautious enterprise buyers. A company letting an AI agent move money or change its systems wants proof that the limits hold even when the agent doesn't follow the rules, and that is exactly the worry you're going after.
Ultimately I'd love the opportunity to have my work truely peer reviewed and published more broadly.
For the peer review you are seeking, a claim-to-evidence map would help a reader choose a bounded starting point: the exact property, assumptions and trusted components, then the artifact or test supporting it and what remains outside scope.
Which single claim would you most like an independent reviewer to try to falsify first? In particular, where do you draw the boundary between enforcing an authorized memory access and establishing that the resulting business action is appropriate? Those seem like distinct review questions.
I have read this post and the exchange, not audited the paper or prototype. I would keep that distinction explicit in any feedback.
@Rhode Brunel Moungondo Thank you for this - genuinely. I've never had anything peer reviewed before, and this is only my third publication on Zenodo, so I'm learning as I go. Your comment was the single most helpful feedback I've gotten so far. You asked the exact questions we needed to answer before this could ever stand up to real scrutiny. Because your critique deserved more than a short comment, I put together a dedicated follow-up post walking through both answers (narrowing our primary falsification target to ODIN-C001 unauthorized-byte preservation, and drawing the PEP vs PDP boundary at ODIN-C012): ai.lounge.ai/p/from-prototype-to-peer-review-answering-rhode-brunel-nov-9-2026 We also published the 14-claim registry, 12 falsification protocols, and opened the repo at github.com/nwalker85/odinos. Grateful for the direction.