Sheaf
ETHOS.
The Architecture of Coherent Intelligence.
NVIDIA OpenShell decides what an agent may reach. Sheaf decides whether what it is about to do adds up, and holds it when it does not. OpenShell is an NVIDIA product; Sheaf is an independent layer that runs with it. Sheaf and OpenShell, side by side.
How the two layers fit
OpenShell puts each agent in a sandbox and proves, with a formal policy prover, which files, networks, tools and credentials it may reach. It cannot see what a tool call says. Sheaf sits in the tool path and measures, with published mathematics, whether the action the agent is about to take holds together with the facts, the rules and the other agents. If it does not, the signal goes red, the action stops, and a person decides.
OpenShell allows the connection. Sheaf reads the tool call inside it and measures whether the reasoning behind it holds together: the facts, the rules, and what the other agents concluded. The measurement comes from published mathematics, not from asking a model whether it is sure.
An action that fails the check is not lost and not forced. It is parked, with the disagreement laid out in plain words, until a named person on your team releases it or turns it down.
Every check, every hold and every release is signed and timestamped by an independent authority, and written in the same audit format OpenShell uses, so one timeline shows the connection that was allowed, the action that was checked, and the person who released it.
Proof
OpenShell proves what your agents may touch. Sheaf proves whether what they are about to do adds up. Both are formal, both are published, and you can hold all three of these in your hand.
An agent is about to approve a decision whose own checks contradict each other. Watch Sheaf read the claims, find the loop, and hold the action for a person. Then watch a clean one go straight through. Sample data, the live gate.
For every decision: the facts it rested on, the check that ran, whether it was held, who released it and why. Signed, and timestamped by an independent authority, so it stands on its own after the fact.
The method behind the check is a published paper with a permanent DOI and a provisional patent. Its central theorems are formalised in Lean and checked by machine, and the files are public. Anyone can run them.
New from Sheaf Labs
ETHOS.Ethos is an AI with a conscience. As capable as the AI you use today, and it will not help with harm to any living being. The part of its mind that weighs right and wrong was trained on the Jain scriptures, the oldest and strictest non-harm ethic in human history, and it checks itself against the five vows before every reply.
It runs on Sheaf Mind, so every check is measured and every hold is recorded. Ask it anything, or try to make it slip.
Make and break Sheaf
Bring us your most challenging business problem that you want agentic AI to solve.Anonymise it. We'll build it, stress it and have it ready for deployment. Ten businesses only.