Two Proofs, One Decision
When an AI system makes a decision that someone later questions, three separate questions get asked. They sound similar in a meeting and they are constantly confused, but they have different answers, different evidence, and different people responsible for them.
Did the organization have a program? Policies, risk assessment, defined roles, human oversight. This is governance.
Was this particular action allowed to happen under the policy in force at that moment? This is authorization.
Did the computation behind the decision actually happen the way you say it did? This is execution.
Most organizations can answer the first with documentation. Almost none can answer the second or the third with anything other than their own logs, written by the party under examination.
Frameworks describe intent. Evidence describes what happened.
An ISO 42001 management system, a NIST AI RMF alignment, a SOC 2 report and a model card all do something real: they establish that a program exists, that controls were designed, and that someone is accountable. That is not paperwork for its own sake. It is the foundation everything else rests on.
But none of them can answer the question an examiner actually asks after something goes wrong, which is always about one specific case. Reproduce what your AI did on this decision, and prove it. Documentation establishes the program. It cannot prove the individual act.
Two different proofs, and they are not interchangeable
We have been comparing notes with IronProof, a Montreal company founded by Dominik Blain that works on the second question. Their approach is formal verification: before deployment, they prove mathematically that the rules implemented in a system match the policy model they are supposed to represent, across the modelled state space. Then, at decision time, an action is evaluated against the policy, authority and transaction state in force, and each approval or refusal leaves a sealed artifact an auditor can re-check months later.
Cyberian works on the third question. Every AI inference produces a cryptographic receipt binding the exact model, the exact input and the exact output. The receipt is issued by a verifier that did not run the job, and anyone can check it afterward, offline, without re-running the work and without trusting the operator. Our architectural rule is absolute: the party that executes is never the party that attests.
The failure modes they address are genuinely different. A policy proof does not detect a quietly substituted model, a silent version change, or an output that was never produced by the model it is attributed to. An execution receipt does not tell you whether the action was permitted in the first place. An AI system making consequential decisions in a regulated environment needs both, and it needs a governance program above them to say why either matters.
Why this belongs to GRC practitioners
Almost every conversation we have about this starts with a governance, risk and compliance professional, not an engineer. That is not a coincidence.
GRC practitioners are the people who get asked to produce evidence, and they are the ones who discover, usually at the worst moment, that what they have is a description of a control rather than proof that the control operated. They are also the people who translate a regulation into something an organization can actually do. When a new form of evidence becomes available, they are the ones who know which obligation it satisfies and which client needs it first.
So the working relationship we look for with advisory firms is simple. You own the governance layer, the management system, the risk assessment, the client relationship. We sit beneath it and supply proof. Third party risk management, in particular, gains something it has never had: instead of asking an AI vendor to describe its controls, an assessor can ask for evidence that the vendor could not have manufactured.
What proof does not do
Being precise about the boundary is part of the point.
A receipt does not tell you the model was the right model for the task. It does not tell you the output was wise, fair or commercially correct. It does not say anything about data nobody disclosed. A formal verification of policy logic does not tell you the policy itself was good policy. Neither replaces professional judgment, and any vendor claiming otherwise is selling something that does not exist.
What both do is remove a specific and increasingly expensive kind of doubt: whether the story an organization tells about its own AI can be checked by someone who has no reason to take its word for it.
Courts have started ordering AI decision records into discovery. Insurance examiners are piloting AI evaluation tools. The EU AI Act now applies. The question is no longer whether organizations will be asked to prove what their AI did. It is what they will hand over when they are.
Cyberian and IronProof are independent companies exploring where their respective evidence layers meet. If you advise regulated clients on AI governance and want to see what a receipt looks like in practice, write to philippe@cyberiansystems.ai.