CoDeS LLCHow governed AI actions work
Separate checks, each with one job
We built the pieces as separate modules, so each can be tested, reviewed and adopted on its own.
ACP, the Action Control Protocol
The permission check: is this exact action allowed?
It ties together who is asking, what, why, where, when and with what limits, as a contract an action must satisfy before it can run. It covers multi-step plans and delegation.
Aegis
The gatekeeper where things cross a boundary.
It checks risk, policy and integrity for requests, tools, documents and other agents. It keeps data from being mistaken for instructions, can deny, hold or quarantine, and records what it saw with tamper-evident hashes.
FES, governed execution
The runner that carries out only approved work.
It stages and tests the work, reports what actually happened, and never approves its own work.
GAIA, the advisor
The advisor that explains and proposes, and holds no authority.
It explains in plain words what is waiting and what is proposed, and carries the decision to whoever holds the authority.
Proof
The receipt of what actually happened.
It is kept apart from all of the above. After an action runs, an independent observation reads back the actual result and compares it with what was approved, and a Proof record keeps the receipt. Blocked work, stale evidence and missing information stay visible as incomplete. They cannot turn into “success.”
UNIFY, the reconciliation step
The step that reconciles the records afterward.
The result is judged on evidence, not on any component’s own report.
Local-first
Authority stays on systems the operator controls.
Cloud access is not the source of authority. Governed workflows are demonstrated on operator-controlled Android and Windows hardware, and are designed to keep working when disconnected.
Who authorizes an action depends on the rules: some actions are authorized by written policy alone, others need a person. The order of the checks depends on the action.
The problem
AI is starting to act, not just answer. As agencies introduce AI assistants that can change files, permissions or workflows, they need to know who approved an action, what the AI actually did, and whether the result was verified. That matters most for high-accountability and intermittently connected systems.
- AI acting outside approved authority
- Limited auditability of autonomous workflows
- Security risk from cloud-dependent AI
- Evidence and chain-of-custody needs
- Operating when disconnected
All of this is research. We are learning which modules fit which partner first.