Patent the Mechanism, Not the Model: How I Protect Mickai

I protect Mickai by patenting the mechanisms that make sovereignty real, not the model that runs inside it. As the founder and named inventor, I chose to file on how the system enforces offline operation, verifiable auditing, hardware-bound licensing and the orchestration of many applications on one engine. Those are the durable engineering inventions. Mickai LTD (Companies House 17166618) holds 104 filed UK patent applications carrying 2,340 claims, patent-pending at the UK Intellectual Property Office. We do not seek protection on any particular AI model, and we do not publish model lineage. The principle is short enough to fit on a business card: we patent the mechanism, not the model.

• The model is a fast-moving, replaceable component; the mechanism that governs it is the lasting invention.
• 104 filed UK applications and 2,340 claims form a moat around the engineering, not the weights.
• We never claim ownership of a model, and we do not disclose its lineage.
• The claims describe how sovereignty, auditability, licensing and orchestration are enforced in practice.
• Everything is filed at the UK IPO and marked patent-pending, stated plainly and without overreach.

Why the model is the wrong thing to protect

Models change every few months. A set of weights that looks state of the art today can be surpassed, retrained or replaced before a patent examiner has finished reading the application. Filing on a specific model would tie the company's protection to the least stable part of the whole system. It would also force disclosure of exactly the thing we treat as an implementation detail: which model runs, how it was trained and where it came from.

There is a deeper point. A model is a component. Mickai is not a model; it is an operating system that can host a capable model, swap it, or run a different one entirely, without changing what makes the product sovereign. Protecting the component would miss the invention. The invention is the machinery around it.

What we actually patent: the mechanisms of sovereignty

The filings describe how the system does what it promises. How inference is kept on the local machine and prevented from reaching the network. How every action is written into an audit record that can be checked after the fact. How a licence is bound to specific hardware so that entitlement cannot simply be copied. How dozens of applications share one engine without leaking data between them, and how an orchestration layer routes work while preserving those boundaries.

In other words, we patent enforcement. Anyone can say their AI is private or auditable. The interesting and protectable question is how that property is guaranteed by the engineering, and made hard to counterfeit. That is what the claims cover: the mechanism by which sovereignty, auditability, licensing and orchestration are actually enforced rather than merely asserted.

104 filings and 2,340 claims as a moat around the engineering

The 104 filed UK applications and their 2,340 claims are deliberately built around the engine rather than the payload. Taken together they map the ways a sovereign system can keep intelligence offline, produce evidence of its own behaviour, and license itself to hardware. The moat is the shape of that map. A competitor is free to pick any capable model they like. What they cannot freely do is reimplement the enforced-sovereignty mechanisms the estate describes.

I should be precise about status, because precision is the point. These are filed applications, patent-pending at the UK Intellectual Property Office. They are not granted patents, and I will not describe them as if they were. Filing establishes priority and puts the mechanisms on record; grant is a separate process that runs on the office's timetable, not ours.

Why patenting the mechanism is honest

Patents demand disclosure. To claim something you must teach it, in enough detail that a skilled reader could build it. That sits awkwardly with any company that would prefer its model provenance stayed vague. By choosing to patent the mechanism, we disclose the part we are comfortable teaching (how sovereignty is enforced) and we decline to make claims about a model we would rather not overstate. It keeps the marketing and the filings in the same honest register.

It also keeps us away from a trap I see often: dressing up a model as an invention. We did not train the laws of probability, and we will not pretend a particular checkpoint is our intellectual property when it is a component we selected. The engineering around it is ours, and that is what we claim.

Why patenting the mechanism is durable

Because mechanisms outlive models. The way a system proves it stayed offline, or binds a licence to a machine, or signs an audit record so tampering shows, remains valuable long after any specific set of weights has been retired. When the model inside Mickai is replaced, the sovereignty machinery is unchanged, and so is its protection. That is the whole advantage of drawing the line where we drew it.

It compounds, too. Each mechanism we harden becomes another candidate for the estate, and each filing is another description of enforced sovereignty that a would-be copyist has to design around. The moat widens as the engineering deepens, not as the model changes.

What being the named inventor means in practice

I am named on these filings because I did the inventing, and I hold myself to the standard that implies. Alongside the applications I have published technical preprints on Zenodo (ORCID 0009-0000-5511-5858) covering the offline attestation record, sovereign inference, private knowledge bases and turning compliance into evidence. The preprints explain the thinking in the open; the filings protect the mechanisms. I would rather be judged on both than on slogans.

It also sets the tone for the company. Mickai is built to be sovereign, auditable and honest about its own limits, and the way we handle intellectual property is part of that. We patent what we genuinely invented, we describe it accurately, and we let the model be what it is: a capable, replaceable part of a system whose value lives in the engineering around it.

Frequently asked questions

Does Mickai own or patent the AI model it runs?

No. We do not seek protection on any particular model, and we do not disclose model lineage. Poros, our inference engine, runs a large capable model locally, but the model is treated as a replaceable component. Our patents cover the mechanisms around it, not the weights inside it.

What exactly do the 104 filings cover?

They cover how sovereignty is enforced: keeping inference offline, producing verifiable audit records, binding licences to hardware, and orchestrating many applications on one engine without leaking data between them. In short, the machinery that makes the system provably private and accountable.

Are these granted patents?

No. They are 104 filed UK patent applications carrying 2,340 claims, patent-pending at the UK Intellectual Property Office. Filing establishes priority and puts the mechanisms on record. Grant is a separate process on the office's timetable, and we will not describe the applications as granted.

Why not just keep the mechanisms secret?

Trade secrecy protects only until someone reverse-engineers or independently discovers the method. A filed application secures priority regardless, and forces us to describe the mechanism clearly, which we consider healthy. We keep model choices private and put the sovereignty mechanisms on record.

Who is the named inventor?

I am. I am Micky Irons, founder of Mickai LTD and the named inventor on the filings. I have also authored technical preprints on Zenodo under ORCID 0009-0000-5511-5858 that explain the underlying ideas in the open.

How does this help a customer?

It means the thing you are buying (enforced sovereignty) is the thing we protect and stand behind, rather than a model that could be swapped out from under you. Your data stays local, your audit trail is verifiable, and the engineering that guarantees both is on record at the UK IPO.