High Assurance AI
High Assurance AI is formally proved infrastructure for delegated execution in high-consequence AI implementations and regulated industries.
Financial services is an industry ready for high-assurance software and AI. Traditional regulation there has focused on people and on a deterministic process, one that gives the same result each time. AI introduces a probabilistic challenge, because a route to the outcome can vary. That challenge lands on the static fundamentals that regulation and process depend on.
What the phrase names
Anthemis uses the phrase for a system given the authority to act where a mistake has a severe cost. A copilot only advises a person. A high-assurance system acts. The engineering sits around the AI Model (AI). The AI is what proposes the next action. Making that AI always give the same result would remove the behaviour people want from it.
The older engineering sense is a system whose failure has a severe consequence. In the aviation supplement DO-333, formal methods are a formal model, a mathematical description of the system, plus an analysis of that description. The proof holds inside that description and the assumptions written down.
Dalrymple and others ask for a safety guarantee that can be measured. It comes from a world model, a description of how the surrounding world behaves, and a statement of what must stay safe. A checker tests that statement. Monitors take over when that description no longer holds. An IETF draft draws the same line. Computation is not authority. An output from the AI is a proposal until an independent check allows the effect. If the check and the effect do not match, the action does not run.
Where Riley Betts puts the work
The AI proposes an action. The institution’s rules decide before anything runs. That is the gate. The Compliance Operator is the current product. It is a policy written so it stays consistent, and a plugin that allows or blocks that proposed action before it runs.
Spec++ is how that policy is written so it can be proved. Lean, a language for mathematical proof, is how it is proved and implemented. Proved means relative to the organisation’s stated limit.
The work is what Riley Betts calls Proof limited AI. The proof includes its limit: what was shown, given which assumptions, and what is still trusted. Peras (πέρας) is the classical word for that limit.
When a third party must see the decision, or must have evidence of it, the receipt is Proof Fabric. Proof Fabric is the gate and Anchor used together. Anchor is a receipt a third party can check, without the private inputs. Anchor is on the active Riley Betts roadmap. Read the Anchor Proof Framework for more information.
The way to start is an authority map for one workflow: what every acceptable action must satisfy, which assumptions that claim depends on, and what is still trusted. Then Lean-proved high-assurance software. The Lean protocol stacks are message protocols written so they can be proved. Anchor is the receipt pattern. Riley Betts has researched and prototyped both. The open-source libraries are the evidence. They are not a second product.
One interesting high-assurance use case is an AI agent confirming an over-the-counter (OTC) trade with another AI agent.
The argument
The question is not whether a system is trustless, with no trust left in it. It is where the trust has gone: what was proved, relative to which assumptions, and what is still trusted. Trust and Proof.
When an AI agent builds the process, assurance cannot mean “the approved process was followed.” The claim is that this process trace meets this goal relative to this boundary. A process trace is the record of what the AI agent did. The boundary is the set of assumptions the proof depends on. The floor is the assurance the institution required, and the boundary has to cover it. This is Goal-Relative Assurance.
An allow, the decision to let the proposed action run, is not a discharge of compliance or trust. The institution, not the AI agent, pins the goal and the floor. Governance is the relation among the goal, the process trace, the proof boundary, and the trust that remains. Agentic Process Governance.