Work
Patents
Protecting the hard parts.
The first provisional was filed on June 5, 2025 — the core execution protocol, the filing everything else references. Twelve more followed, and the core has since been extended into an international filing family.
Together they cover the three questions that appear the moment an AI agent takes real actions:
Authorization — “who says the agent can do that?”
Reliability — “does the agent actually work?”
Accountability — “what did the agent actually do?”
What the portfolio covers — as ideas
the anchor
The execution protocol
A machine-native way for software to transact: one structured message in, one terminal outcome out — execute, or instruct on exactly what is missing. Never a dead-end rejection.
language → structure
Intent compilation
Natural language becomes a structured, checkable program before it is allowed to become an action.
two kinds of computing
The boundary
A hard line between probabilistic computation and deterministic enforcement — where a model's suggestion ends and a system's commitment begins.
no model in the loop
The firewall
No language model in the authorization path. A prompt injection can talk to the model all day; it cannot reach money.
propose ≠ commit
The commit split
The authority to propose and the authority to commit are held apart — with holds, locks, and quotes that survive until a real decision is made.
no partial actions
Completeness before execution
An action either has everything it needs, or the system states precisely what is missing — before anything runs.
receipts that outlive systems
Proof of execution
Signed, verifiable evidence of what happened — and of what was refused — checkable offline, long after the systems that produced it are gone.
state machines, not heroics
Autonomous recovery
When execution breaks mid-flight, the system returns itself to a known-good state — deterministically, without a human untangling it.
delegation chains
Agent-to-agent integrity
When agents delegate to agents: authority that can only narrow, never widen — and revocation that cascades through the whole chain.
safety at wire speed
The fast path
Making the safe path the fastest path — precomputed, algebraic execution so governance costs microseconds, not minutes.
compounding
Network intelligence
Execution that learns: every transaction on the network makes the next one safer and cheaper.
Why only concepts
This page is as specific as it gets, on purpose. The ideas are public; the mechanisms stay in the filings until they publish on their own schedule. When a concept is filed and worth explaining, it becomes a whitepaper — the argument, never the internals.
The public face of these ideas
The Authorization Chain
Authorization is not a check. It is a chain.
The Action Layer
Models generate. Agents decide. Protocols route. None of them control what is allowed to happen.