Allodial — What we add to what you already sell

What is the gap on your shelf?

We have that layer and no UK paper. You have the paper and an empty shelf where that layer would sit.

You sell computers to British institutions. You have the accreditations, the frameworks and the customer relationships. What you almost certainly do not have — because we could not find it offered anywhere in this channel — is a way to tell a buyer that a decision the machine produced was correctly made.

None

Correctness, assurance or formal-verification claims about a decision that we could find offered anywhere in the UK public-sector compute channel.

Our own survey
£207,720

The 2026 sub-central contracting threshold, including VAT. One unit of our entry tier prices inside it.

PPN 023
50–60°C

Delivered liquid, at the top of the Greater London Authority's own heat-recoverability scale.

Greater London Authority

About revision 2

Revision 2 restructures rev1. The value proposition now precedes anything defensive; the warranty is stated as the competitive claim it is rather than as a disclaimer; and diligence items are compressed into one section at the back. Substance is unchanged — nothing has been softened or removed, only reordered and put in its proper register.

We are not asking you to change what you sell. We are adding a line item you cannot currently offer.

We read the published claims of every significant UK public-sector reseller, integrator and HPC specialist we could identify, looking for a correctness, assurance, auditability or formal-verification claim about a decision. We found none in that channel. Its assurance vocabulary describes where the data sits and how the supplier is run — not whether the answer is right.

A warranty a vendor can actually stand behind is one an insurer will price; a warranty over "model behaviour" is one nobody can.

What we make

A small liquid-cooled inference appliance, with a proof harness in front of every decision that matters.

2 kW to 70 kW, one chassis, running an open-weight model with a Lean 4 proof harness. It sits inside the customer's building, on their own power, and delivers 50–60°C liquid to whatever wants heat.

Three things follow from that form factor, and each is a sale on its own

It sits under every regulatory line being drawn. A single unit is two to four orders of magnitude below the thresholds in current data-centre legislation. No rezoning hearing, no developer, no undisclosed end user, no planning fight. For a buyer watching what happens to large data-centre applications in their own borough, that is the whole conversation.
The customer keeps control, provably. Open weights buy verifiable provenance, a weight hash they can pin, no licence server, no activation, no phone-home, and no vendor who can switch the machine off. Sensitive data is encrypted at rest and in transit under customer-held keys; the appliance runs default-deny at the network boundary, with every permitted flow enumerated, logged and revocable; vendor access is customer-initiated, time-boxed, session-recorded and attested, with the recordings in their custody. This is a controlled boundary the customer owns — which is a better product than an air gap and a great deal easier to support.
The heat is not a rounding error to the right host. 50–60°C delivered liquid sits at the top of the Greater London Authority's own recoverability scale, which rates liquid-to-liquid at 30–60°C as "very high suitability" against air cooling at 15–25°C as "low-moderate". Because the appliance sits inside the building that wants the heat, there is no network to join and no heat pump in the chain.

For an institution of typical British scale the answer is usually one unit, and which tier depends on the functions in scope. We size from the counts a buyer already publishes, never from population.

What exactly does the harness prove?

Lean protects the artifact, not the advice.

The harness proves that a stated conclusion follows from an encoded rule set applied to structured inputs. On a benefits determination, a permit screen, a records redaction or a utility bill, that means:

The conclusion is the correct application of the written rule, with the clause cited
Arithmetic over a rate table, fee schedule or benefit formula is exact
Every value that fed the decision traces to a named field in a named record
The system refused rather than answered when inputs fell outside the proved domain

The warranty, which is narrow on purpose

We warrant that the specified checkers executed, that the enforced policy matched the stated policy, and that the artifact hash and transcript are complete and unaltered.

What the warranty covers, and what it does not.
What we warrant	What we do not
The specified checkers executed.	Model behaviour.
The enforced policy matched the stated policy.	The advice, as opposed to the artifact.
The artifact hash and transcript are complete and unaltered.	The whole workload — most of it is prose and falls outside the harness.

That narrowness is the product. Every competitor claiming coverage of everything is claiming coverage of nothing, and sophisticated buyers know it. A warranty a vendor can actually stand behind is one an insurer will price; a warranty over "model behaviour" is one nobody can. We say precisely what we cover, which is why the cover means something.

The same logic runs through how we report. Coverage percentage × accuracy percentage, measured monthly — never "the system is correct." Coverage is a minority of volume and a majority of consequence: most of a real workload is prose and falls outside the harness, and what the harness covers is the fraction where something irreversible happens — payments, eligibility determinations, records that get written.

We price and pitch on that fraction, and saying so is what makes the number believable.

Does this compete with what you already carry?

Why this is additive rather than competitive.

We read the published claims of every significant UK public-sector reseller, integrator and HPC specialist we could identify, looking for a correctness, assurance, auditability or formal-verification claim about a decision. We found none in that channel. Its assurance vocabulary is:

ISO certification
Cyber Essentials
NCSC accreditation
CREST
"Sovereign"
UK data centres

All of which describe where the data sits and how the supplier is run. None of them describes whether the answer is right.

To be precise: formal verification applied to software is a real and well-funded field, and there are serious companies in it. Our claim is narrower and checkable — this channel does not offer it as a line item, and the buyers' own policy framework already asks for something very like it.

The expectation is already written down, in two instruments of different force

Ethics, Transparency and Accountability Framework for Automated Decision-Making

Cabinet Office guidance rather than statutory duty. It asks public bodies to "ensure that your algorithm or system adheres to the necessary legislation" and to "institute formal review points (recommended at least quarterly)."

Algorithmic Transparency Recording Standard

The harder of the two: mandatory for ministerial and non-ministerial departments and public-facing arm's-length bodies.

Today both are discharged with documents and review boards. We could find nobody in this market selling a tool for either.

How hard is this to actually sell?

Two commercial facts that make this easier than it sounds.

It prices inside the sub-central threshold.

The 2026 threshold for sub-central contracting authorities — which is what universities are — is £207,720 including VAT per PPN 023. One unit of our entry tier prices inside that at any plausible exchange rate, and still does over a five-year support term, which takes it out of the advertised-competition regime under the Procurement Act 2023.
The buyer that fits first is technically self-sufficient.

University research-computing staff and central-government digital teams can evaluate a formal-verification stack on its merits. That shortens every conversation, and it is where we would start.

What are we actually asking for?

Not a term sheet. A view on three questions.

Would your buyer value a claim about decision correctness?

And if not, what would they value instead?
Can you deliver and support a liquid-cooled chassis at this density?

And what would you need from us to do it?
What shape of arrangement is worth your while?

Line item, reseller agreement, or something we have not thought of.

We would rather hear your structure than propose ours.

What would you find out later?

The questions you would ask on the first call, answered now.

We would rather you had these from us than found them later. None of them is a surprise to us and none should be a surprise to you.

Reference customers. We are pre-revenue and no unit has been deployed. Anyone carrying us is carrying a first-of-kind.
Who has bought this shape of thing. Universities and research councils buy on-premises GPU compute. We could find no instance of a UK local authority, or a US city or county, doing so. That is either a large gap or a large warning and we think it is partly both — which is why we would lead with higher education and central government and treat councils as the later market.
Where our sizing comes from. The capacity model is built on US municipal workloads. The UK equivalents are an extrapolation we would want to test with you rather than assert, and NHS trust and university-department workloads are not yet modelled at all.
What the proof does not reach. A proof covers the step from typed fields to determination. It does not cover the step from document to typed fields — that is perception, and no proof assistant reaches the world. What covers that is traceability to a named source field, a halt to a human caseworker on ambiguity, and measured extraction accuracy reported alongside coverage.
What is not built yet. Cross-jurisdiction capacity sharing is designed and on our roadmap, not shipping. The full network-flow inventory — every direction, destination, port, protocol and purpose — is in progress and not finished.
Export control. Classification of the software layer is with counsel and not settled. We will not put a technical package or a distribution commitment in front of a partner until it is.
And on the threshold point above. Pricing inside £207,720 is not the same as a direct award, and we will not pretend it is. Your institution's own standing financial instructions will set a far lower competitive-quote threshold, and we have not yet had UK counsel confirm the current figure, on which we have seen one conflicting published number. Treat it as "this does not have to be a full tender", not as "this can be bought without process."

What we add to what you already sell.