Public Sector AI and the Decision a Citizen Can Challenge

By Micky Irons, founder of Mickai.

Justice in public administration is not a sentiment. It is the practical ability to weigh a decision and see how it was reached.

A benefits claim is refused. A visa is flagged. A child is moved up a risk register. Increasingly, the first pass on these decisions is made by a model, and the human in the loop signs what the model proposed. The public conversation fixates on whether the model is accurate. That is the wrong test. The legitimate test for any state decision is whether the citizen affected by it can challenge it, and whether the answer the state gives holds up when they do.

Accuracy is a statistical property of a system across a population. Accountability is a property of a single decision, owed to a single person. A model can be ninety-five per cent accurate and still owe an unanswerable account to the one citizen in twenty it got wrong. Public administration runs on that second standard, the right to reasons and the right to redress, and most deployed public sector AI cannot meet it.

Why most government AI cannot be challenged

The failure is rarely the model. It is the record around the model. When a decision is contested months later, three things tend to be missing. The exact version of the model that ran is gone, quietly replaced by a retrained successor. The specific inputs that produced this output were never captured, only logged in aggregate. And the log itself sits in a database that the same authority can edit, so it proves nothing to an outside party who already suspects the authority.

This is why so many algorithmic decisions collapse the moment a court or an ombudsman asks the obvious question: show me, for this person, on this date, what the system saw and what it did. An answer that the deciding party could have altered after the fact is not evidence. It is an assertion. The citizen is asked to trust the very institution they are disputing.

Memory that the deciding party can rewrite is not a record. The seal has to come from outside the institution being questioned.

The decision a citizen can challenge has three properties

Strip the problem to its requirements. A decision is genuinely challengeable only if it is reproducible, explainable, and tamper evident. Reproducible: anyone can re-run the exact model version against the exact inputs and get the same output. Explainable: the reasons are recorded at the moment of decision, in terms the citizen and their advocate can read, not reconstructed afterwards. Tamper evident: the record carries proof that nobody, including the issuing authority, altered it after it was sealed.

Most govtech procurement asks for the first two and forgets the third, which is the only one that matters when trust has already broken down. Reproducibility and explanation are worthless if the authority can quietly revise the inputs before producing them. Tamper evidence is the property that lets a citizen win an argument against the state without the state's cooperation.

Sovereignty is the precondition, not the feature

There is a second, quieter failure. Public bodies handle the most sensitive data a person has, and a great deal of public sector AI ships that data to a commercial cloud and an external model provider to get an answer back. The citizen never consented to that journey, and the authority often cannot say precisely where the data went or what was retained. A decision made by sending a vulnerable person's file to a third party's servers is compromised before the model even runs.

Sovereignty here means something concrete: the decision is computed on hardware the public body controls, with no dependency on an outside provider to function. It is the precondition for everything else, because a record you do not hold is a record someone else can change or withhold. Mickai is built around exactly this constraint. It is a Sovereign Intelligence Operating System, a SIOS that runs fifty specialised AI brains (twenty-five domain and twenty-five operational) on the operator's own hardware, fully offline-capable. The data does not leave. The decision is made where the authority can stand behind it.

Sovereignty is the burden a public body has to carry itself. A decision computed on someone else's servers is a decision it cannot fully answer for.

Sealing the decision so it survives the challenge

Running the model in house solves where. It does not yet solve whether the record can be trusted later. For that, Mickai writes every consequential action to the Open Audit Record, the OAR. When a decision is made, the model version, the inputs, the output, and the reasons are captured and then sealed and signed with FIPS 204 ML-DSA-65, the published NIST post-quantum signature standard. Mickai did not invent the standard. It adopts it, which is the point: the proof rests on a public cryptographic norm, not on a vendor's word.

A signature proves the record was not altered after sealing. It does not, on its own, prove when the record existed, and a sufficiently determined authority could in principle regenerate a whole sealed history. That is what Pantheon closes. Pantheon is Mickai's own sovereign Layer 1, anchored to Bitcoin. It takes a hash commitment of the record and anchors that single fingerprint to Bitcoin, fixing the record in time against the most expensive clock to forge that exists. It does not move any Bitcoin and it is not a Bitcoin Layer 2. Anchoring is not spending. The citizen, the ombudsman, or the court can later confirm that the record produced today is the same record that existed on the day the decision was made.

What this changes for the citizen

Put the three layers together and the citizen's position is transformed. The decision was computed on the authority's own hardware, so the data never left. It was sealed at the moment it was made with a public post-quantum standard, so it cannot be quietly rewritten. Its fingerprint was anchored to Bitcoin, so its date cannot be backdated. The citizen no longer has to trust the institution. They can verify it, and so can anyone they bring with them.

This is not a hypothetical research posture. The approach is backed by 101 filed UK patent applications, around 2,234 claims, owned by Mickai LTD, with the named inventor Micky Irons. The detail worth holding onto is simpler than the cryptography. A public sector decision is only legitimate if the person on the receiving end can challenge it and the answer holds. Build the record so it holds, and the rest of the accountability conversation finally has something solid to stand on.