104 UK patent applications. 2340 claims. One named inventor.

Anyone building sovereign or on-premise AI eventually reaches the same conclusion Mickai reached first: a single model cannot be trusted to authorise a consequential action on its own, so the safe architecture is many independent brains that must agree before anything dispatches. The moment a competitor adds a second model to cross-check the first, gates a sensitive action on their agreement, refuses on divergence to catch hallucination, binds authorisation to hardware-attested keys, or partitions brains into isolated department fleets on an owned offline box, they are standing on this filed ground, because these filings claim the cooperative substrate, the signed envelope bus, byte-equivalent cross-model quorum, divergence-triggered refusal, replay-resistant voice-to-action binding, TPM-attested quorum with automatic rollback, and department-partitioned offline orchestration as an integrated whole rather than one narrow trick. The filings also pre-empt the obvious escape routes by distinguishing cloud multi-tenant isolation, single-tenant inference frameworks, and confidential-computing enclaves, so the design space around a sovereign, offline, consensus-gated multi-brain system is already occupied, which is why the sensible next step is an inbound licensing conversation rather than a redesign.

If your roadmap has more than one model checking another before a sensitive action fires, you are building on ground Mickai has already filed, and a licence is faster and safer than a redesign.

Anyone building sovereign or on-premise AI has to solve the inbound perimeter, and this family has already filed the ground on which that perimeter is built. A credible sovereign system cannot let raw requests reach a model unclassified, cannot allow retrieved context to carry hidden instructions, and cannot answer a below-clearance query in a way that betrays that classified material exists, yet those are precisely the behaviours the filed inventions claim, right down to gating access per voiceprint rather than per login and making revocation reach backwards through the audit record. The pieces are not severable, because the value comes from binding classification, injection containment, clearance-ceiling retrieval, and per-voiceprint revocation into one boundary, so a competitor who reimplements any serious version of it lands inside filed claims. Filed applications establish priority dates and form a published prior-art record, which is why the sensible path for a serious builder is an inbound licensing conversation rather than a redesign that still reads onto the same ground.

If you are building a sovereign or on-premise AI perimeter, the classification, injection containment, and clearance-gated retrieval you need is already filed here, so talk to us about a licence before you build onto it.

Any organisation building sovereign or on-premise AI eventually has to answer one question from its regulator: prove what the system did, on keys you control, in a way that survives both a vendor's disappearance and the arrival of quantum computing. The moment a competitor tries to answer it properly they reconstruct this family, a post-quantum signed ledger, a causally linked lineage graph that walks back to the originating prompt, a vendor-neutral record format, an offline browser-resident verifier and a separated audit-witness domain, because those are the only architecture that actually satisfies an independent auditor rather than merely logging events. These are filed UK applications with established priority dates, part of 104 filed applications and 2,340 claims owned by Mickai LTD, and they sit precisely where the verification, cryptography and trust-separation ground converge. A serious buyer of sovereign AI does not route around this; they find they need a licence to it, which is where the conversation with us begins.

If your sovereign AI has to prove to a regulator what it did, on your own keys and in a form that survives quantum computing, this is the filed ground you will need to license.

Any organisation building sovereign or on-premise AI that is allowed to take consequential actions, moving money, deleting records, signing contracts, commanding hardware, must solve the same problem this family already occupies: how to prove a specific living person authorised a specific action, how to undo that action cleanly if it was wrong, and how to be sure of the effect before it commits. The filed inventions do not describe one feature; they describe the interlocking substrate, a typed-action ontology whose every action carries a declared inverse, deterministic voice gating bound to that ontology, dry-run simulation over the same typed actions, compensating rollback over the same declared inverses, and forecast-driven thresholds that tune the gates, so a competitor who adopts even a fraction of the pattern is reading directly onto claims with a UK priority date. This is the ground a diligence lawyer or a security auditor will insist a serious sovereign-AI product stands on, which is precisely why the sensible move is an inbound licensing conversation rather than an attempt to invent around six mutually reinforcing filings at once.

If your on-premise AI is permitted to act on the real world, the authorisation, simulation and reversible-undo layer it needs is already filed here, so the efficient path is to license it.

Anyone building sovereign or on-premise AI that lets an agent act, and not merely answer, has to cross this filed ground. The moment an autonomous agent can write a file, run a command, place an order, rewrite its own skills or edit an image, the operator needs an enforced boundary the agent cannot bypass and an offline record that proves what the boundary did, which is precisely the interceptor, the sealed plan-execute-verify ledger, the pre-inference model attestation, the hardware-attested skill-mutation gate, the verifiable reasoning ledger, the policy-shadowing diff and the consent-gated refusal set out here. There is no partial route. Bolting logging on after the fact does not deliver structural prevention, and running the gate inside the agent's own process does not deliver independent proof, so a competitor either builds the perimeter as its own trust domain with post-quantum signed evidence, in which case it reads onto these applications, or it ships an agent that can fail destructively with nothing to show a regulator. That is why the sensible move is an inbound licensing conversation rather than a redesign attempt.

If your roadmap includes an AI agent that takes real actions or improves itself, it is faster to license the perimeter that makes destructive failure impossible by construction than to rebuild it around filed claims.

Anyone building a sovereign or on-premise generative system faces the same non-negotiable customer demand: prove what your model produced, prove nobody altered it, and prove it offline without trusting the vendor's servers. Mickai has filed on the mechanism that satisfies that demand at the level of the individual generated element across every modality a buyer will ship, imagery, video, live voice, cloned voice, avatars, music, code, translation, design systems, and game worlds, each seal bound to a key the operator holds in their own hardware and each output walkable back to its originating prompt. A competitor can reach for container-level or file-level watermarking, but the moment a regulator, a court, or an enterprise procurement team asks for per-frame, per-pixel, per-token, or per-line attestation that survives editing, transcoding, and re-encoding, they land on this filed ground. Because these applications are published and establish priority dates, they form a prior-art record that a rival cannot quietly reinvent, which is precisely why a serious sovereign-AI builder ends up talking to Mickai rather than around it.

If your sovereign or on-premise system generates anything a regulator, court, or customer may later need to authenticate, this filed family is the ground you build on, so let us open a licensing conversation.

Anyone building sovereign or on-premise AI that composes or advises, rather than merely chats, has to touch this ground. The moment a system lets a user edit a document, composite an image, revise a model, cut a video, generate a game, run a scenario, or issue a recommendation, the regulated customer needs to prove after the fact how the result was reached, whether it can be undone, whether it reproduces, and what limits governed the advice. Mickai has filed the primitives that make each of those provable at the granularity of the individual edit and run: type-safe inversion for retroactive section-level undo, source-anchored transformation lineage, deterministic seed-sealed simulation, and disclaimer binding that survives forwarding. A competitor can rebuild a generative feature, but the sealed, replayable, reversible, disclaimer-bound record around it is the part their own auditors and regulators will demand, and that record is exactly what these filed applications describe as prior art. That is why the conversation comes inbound: the capability is easy to ship and the accountable version of it is already on the register in Mickai's name.

If your sovereign AI edits, composites, simulates, or advises and has to prove it afterwards, the accountable version of that capability is already filed here, and the shortest path to shipping it is a licence.

Anyone building sovereign or on-premise AI has to answer a single question, who owns the trust anchor, and every honest answer walks straight into this filed ground. If the chip's identity is set at fabrication or held by a vendor, it is not sovereign; the moment a designer moves that identity to the operator and burns it at first power-on, they are inside the operator-personalised silicon root of trust that Mickai filed. The neighbouring claims close the obvious escapes: making the accelerator removable invites the host-acceptance attestation inversion, letting operator state follow the silicon invites the on-silicon migration and accumulating audit chain, cutting the vendor out of updates invites the operator-controlled bootstrap with air-gap mode, and giving customers verifiable model names invites the signed alias layer that binds identifier to weights, corpus, and recipe. These are filed UK applications, so they already establish priority dates and form a published prior-art record that a competitor's later filings must contend with, which is precisely why the sensible move is an inbound licensing conversation rather than a design-around that keeps colliding with the same five inventions.

If your sovereign accelerator personalises its identity to the operator, survives a host swap, or lets a customer verify a model name against its weights offline, that ground is filed here, so talk to us before you build it.

Anyone building sovereign or on-premise AI ends up making the same promise this family already occupies: that the system runs on hardware the customer owns, offline, air-gapped when required, and that the offline state is provable rather than asserted. Every serious buyer in defence, nuclear, intelligence and regulated finance now asks for exactly that proof, and the honest ways to deliver it all land inside filed Mickai ground: continuous hardware-emitted attestation of interface and process state, pre-loaded audit anchors that survive months of disconnection, capability gating and energy-aware scheduling sealed into the same chain, and a continuity demonstrator verifiable from the device public key with no network, no certificate authority and no vendor in the trust path. The standard remote-attestation, tamper-evident-logging, measured-boot and time-stamping mechanisms all require a connected verifier, which is precisely the thing an air-gapped deployment cannot have, so they cannot reproduce the property without reading onto these applications. That corner is filed and forms part of the published prior-art record, which is why a competitor reaching this ground has a reason to open a licensing conversation with us rather than around us.

If your roadmap requires proving to a regulator that a device stayed offline, that proof runs through filed Mickai ground, and the right next step is a licensing conversation with us.

Any organisation building sovereign or on-premise AI eventually confronts the same three obligations this family covers, because they are not features but legal duties: a memory that outlives its owner has to pass to someone, and a regulated system has to be able to prove it forgot. Once you accept that inheritance and erasure must be provable rather than asserted, you are on filed ground. Mickai's applications claim sealed post-mortem activation, signed inheritance write-back to a successor identity, clearance-gated heir bequests of memory and personal archives, and cryptographic proof of both dataset and model-weight erasure under Article 17, all bound to operator-held silicon and one post-quantum audit chain. A competitor can build a delete button, but it cannot build a verifiable proof of erasure or a revocable, auditable inheritance without crossing claims already on the public register, owned by Mickai LTD. That is precisely the point at which the sensible move is an inbound licensing conversation rather than a redesign.

If your roadmap includes digital estates, successor identities or a defensible answer to a right-to-erasure request, this filed family is the ground you will need to license.

Anyone building an on-premise or sovereign trading system that wants both low decision-to-order latency and independent provability lands on this filed ground by necessity. The obvious sovereign design, keep the model and the router on hardware you own so nothing crosses a vendor boundary, is the exact colocation the family claims, and the obvious way to make such a system defensible to a regulator, seal each order under an operator key and publish a backtest that reproduces from public data, is the exact attestation and reproducibility the family claims. The cooperative ensemble with a tail-risk veto, the liquidity-aware cascade, the voice-gated and bounded execution, and the signed trade lineage then cover the practical mechanisms a serious desk needs to run and to prove what it ran. These 104 filed UK applications, 2,340 claims in total, form a published prior-art record with established priority dates, so a competitor cannot quietly ship the same architecture and cannot easily invent around a moat that mirrors the only sensible way to build the thing. That is precisely the position that turns a build decision into an inbound licensing conversation.

If your sovereign trading roadmap puts the model and the order router on the same owned box and promises regulators a provable audit trail, that architecture is already filed here, so the efficient path is to license it rather than design around a published priority record.

Anyone building sovereign or on-premise AI that touches more than one device, or that reaches into the physical world through actuators, lands on this filed ground the moment they decide the owner, not a cloud control plane, holds authority. The obvious architectures all re-centralise: a fleet needs a coordinator, a robot needs a controller, and the path of least resistance is to make that coordinator the trust root, which is precisely the master-node model this family displaces by making the operator's key the master and every device and every actuator a signer beneath it. Once motion must be attributable for liability, insurance, or regulatory audit, the per-actuator signing of physical actions and the 10 Hz mode-attestation taxonomy become the natural, and now published and priority-dated, way to answer who was in control, so an embodied-AI or fleet builder finds the defensible design space already occupied. Filed as part of 104 UK patent applications carrying 2,340 claims and owned by Mickai LTD, this record is what a competitor's own patent counsel surfaces during clearance, which is what turns a design review into an inbound licensing conversation.

If your fleet or your robot answers to a master node instead of the owner's key, or its actuators move without signing, this filed family is the ground to license before you ship.

Anyone building sovereign or on-premise AI eventually needs its actions to be more than locally logged, they need external witness, cross-organisation settlement, continuous regulator-verifiable compliance and continuity that survives key rotation and the loss of the operator. Every credible route to those properties runs straight through this filed ground, because the inventions do not claim a generic ledger, they claim the specific rule that consensus, settlement, anchoring, supply locking, the safety gate, the bridge and validator succession are all conditioned on records sealed before admission under an operator-held post-quantum key. A rival cannot reach the same assurance by attesting actions after consensus, by anchoring without pre-sealing, or by hardening the bridge and the key migration separately, because the sealed-before-admission structure is the invention rather than an implementation choice. That is why a serious sovereign-AI or chain team reading this file tends to open an inbound conversation, the cleanest path forward is a licence rather than a decade of design-around.

If your sovereign or on-premise AI needs its actions witnessed, settled, continuously compliance-verified and able to survive incapacity, this filed family is the ground to license rather than rebuild.