Sovereign AI for Courts and the Judiciary

A court is one of the few institutions where the record is the outcome. A ruling that cannot be traced back through its reasoning is not a lighter version of justice, it is a different thing entirely. So when artificial intelligence arrives at the door of the judiciary offering to summarise bundles, surface precedent and triage caseloads, the first question is not whether it is clever. The first question is whether it can be trusted with privilege, and whether anyone can prove, months or years later, exactly what it did and why.

That is the problem we set out to answer. Mickai is a Sovereign Intelligence Operating System, a SIOS, and we designed it for exactly this boundary: the place where sensitive material cannot leave the building, where every action must be accountable, and where the technology must earn a seat rather than assume one. Case intelligence with privilege intact, and signed provenance for every output, is not a feature we bolted on. It is the reason the system exists.

The judge in the underworld and the weight of the record

In myth, Rhadamanthus judged the dead. He was chosen not for mercy or for eloquence but because he was incorruptible and because his judgements were remembered. Every soul that passed before him received a reckoning that could not be quietly rewritten afterwards. That is a surprisingly exact picture of what a modern court needs from its tools: not a machine that decides, but a machine whose every contribution to a decision is permanent, attributable and open to challenge.

Most court technology fails this test in a mundane way. A model produces a summary, a clerk relies on it, and six months later nobody can reconstruct which documents fed that summary, which version of the model produced it, or whether privileged material leaked into a response it should never have touched. We treat that gap as unacceptable. In our design, the record of what the system did is as durable as the ruling it helped inform.

Privilege that is enforced, not promised

Legal professional privilege is not a preference to be respected on a best-effort basis. It is a legal boundary, and in a court setting it sits alongside data-protection duties under the General Data Protection Regulation (GDPR) and, for cross-border and defence-adjacent matter, export controls such as the International Traffic in Arms Regulations (ITAR). Any system that sends bundles to a shared cloud to be processed has, by construction, moved privileged material outside the walls where privilege is guaranteed.

Mickai runs on hardware the customer owns, air-gapped or on-premise, with zero data egress. Nothing about a case leaves the estate. The brains that read a bundle, cluster related exhibits or draft a chronology operate inside that boundary, and privileged content never becomes training data for anyone. Where a matter demands stricter separation still, brains can be scoped so that material from one case is cryptographically walled off from another, and any brain can be revoked instantly if its access should end.

Signed provenance for every decision-support output

This is the heart of it. Before any brain acts, whether it retrieves a document, drafts a case summary or flags a limitation-period risk, it writes an Operation Attestation Record, an OAR, that describes the action and is signed before that action executes. The signature uses post-quantum cryptography, specifically the FIPS 204 ML-DSA-65 standard, so the attestation remains sound even against future computing threats that would break today's signatures.

Those records are linked into a tamper-evident chain using SHA-3-512 hashing, so the sequence of what happened cannot be reordered or edited without the break showing. The result is a cryptographically-signed audit ledger that a judge, an opposing party or an appellate court can verify offline, on their own equipment, without trusting us and without a network connection. A decision-support output is no longer a floating claim. It arrives with a proof of exactly which inputs, which brain and which version produced it, and that proof travels with the case.

Case intelligence that respects the shape of a case

Judges and clerks do not need a chatbot. They need the workload to shrink without the responsibility shifting. Inside the boundary, our brains do the patient, high-volume work that consumes court time: building chronologies from disclosure, surfacing the authorities that actually bear on an issue, cross-referencing witness statements for inconsistency, and preparing neutral summaries that a human then reads, tests and owns. The system proposes; the judge disposes. That line is never blurred.

For genuinely high-stakes steps, the guardrails tighten automatically. An action that would, for example, release a redacted bundle or finalise a sensitive filing can be configured to require multi-brain agreement plus a voice-biometric approval from an authorised officer. No single brain, and no single point of failure, can push an irreversible step through on its own. The court keeps the kind of layered human control it already trusts in its physical processes.

Built for the regulators who will ask

A court that adopts intelligence tooling will answer to more than its own conscience. The European Union Artificial Intelligence Act (EU AI Act) places justice-sector systems in its highest-scrutiny tier, demanding logging, human oversight and traceability. Emerging standards such as the ISO 42001 management standard for artificial intelligence and the National Institute of Standards and Technology's AI Risk Management Framework (NIST AI RMF) ask for exactly the kind of documented, testable governance that a signed ledger provides by default.

Because provenance is generated at the moment of action rather than reconstructed afterwards, compliance stops being a separate reporting exercise. The evidence that the system behaved lawfully is the same evidence the court relies on operationally. That alignment is deliberate. We would rather the accountable record be a natural by-product of the work than a document someone has to assemble under deadline.

Allies at a different layer

None of this is a quarrel with the public cloud. The large providers, OpenAI, Microsoft, Amazon Web Services, Google and Oracle, do extraordinary things at a scale no court needs to replicate. They operate at a different layer to the one a judiciary must protect. Mickai sits at the sovereign boundary the public cloud, by its nature, cannot cross on the customer's own terms: inside the estate, under the court's control, leaving nothing behind. The two are complementary, not competitors.

The capabilities described here are contained in a body of 104 filed United Kingdom patent applications, comprising about 2,340 claims and owned by Mickai LTD. We frame that portfolio by what it protects: the attestation-before-execution model, the hash-linked ledger and the boundary-preserving architecture that make sovereign case intelligence possible in the first place.

The bottom line

A court cannot adopt a tool it cannot audit, and it must never adopt one that quietly relocates privilege. Mickai answers both. It keeps case material inside walls the court controls, it lets brains do the heavy, repeatable work without ever making the decision, and it signs the provenance of every output so that what the system contributed can be verified, offline, for as long as the record must last. Rhadamanthus was trusted because his judgements were remembered exactly. That is the standard we hold ourselves to, and it is the only standard the judiciary should accept.