Plausible Is Not the Same as True

Plausible is not the same as true. And in the work that matters, the difference is everything.

Earlier pieces on this blog were about seeing: how a third of the overseas companies that own British property still will not tell you who is behind them, and how the owner the register hides is rarely hidden everywhere at once. This one is about what that seeing actually turns up, and what you can build on top of it.

Start with the seeing, because it is the foundation for everything after it.

Most AI today is fluent and ungrounded. Ask it who owns a building, whether a claim on a website is true, whether a company is what it says it is, and it will give you a confident, well written answer. Sometimes that answer is right. It has no way of knowing which times, and neither do you. It is producing something plausible. It was never producing something true. In casual use that is a nuisance. In compliance, due diligence, fraud, investigation and government, it is a liability, because a plausible guess about who owns a company or whether money is clean is not a smaller version of the right answer. It is a different kind of thing.

So we built something that does not guess. Kronaxis Intelligence reads the whole public record against itself, at the scale of the entire country rather than one filing at a time. Companies House, the property register, charges and lenders, planning, the Gazette, litigation, sanctions lists, insolvency, the whole paper trail a modern economy leaves behind. Each of those was built to be read on its own, one record at a time, and read that way each says almost nothing. Read against each other, the truth surfaces in the pattern.

Here is the kind of thing that falls out when you do.

A director a court has banned from running companies, still quietly in control of live ones, because the ban sits in one register and the control sits in another and nobody was reading them together.

People who died years ago, still listed as the controlling owner of active companies, the business carried on in a dead person's name because a dead owner asks no questions and signs whatever is put in front of them.

A sanctioned name sitting behind a chain of shells that the state itself has already registered, the very list designed to catch them looking straight past them because the connection was one hop further out than anyone checked.

An ownership chain that climbs through a dozen British companies, each one owning the next, and only resolves when it finally surfaces in an offshore fund at the top, so that the thing you actually deal with in this country is owned by something you were never meant to be able to name.

And the property owners the register was built to reveal and quietly does not: around thirty thousand titles, out of more than ninety thousand held through overseas companies, that name a trust, another company, or nobody at all, clustered in exactly the cases most worth looking at. That is the strand Private Eye took and ran.

Now the important part, because ownership is only one of the questions the same graph answers.

Point it at a company's own website and it will tell you whether the site is telling the truth: whether the registration number it displays really belongs to it, whether the entity behind it exists at all, whether the reassuring claims match the record or quietly contradict it. Point it at a name and it resolves the network in both directions at once, the people and companies that name controls and the ones that quietly control it, then lays adverse media, sanctions and watch lists, disqualifications and insolvencies over the top, so the risk that standard screening slides past surfaces where it actually sits rather than where a single list happens to look. Point it at public money and it shows where contracts, grants and subsidies concentrate, and around whom. And the same discipline that reads the record reads the open web itself, scoring the authority, the character and the truthfulness of what is published, because in a world where any amount of convincing text can be generated on demand, knowing whether a source is credible is the same problem as knowing whether an owner is real.

It is not a database you query. It is a resolved picture of who and what, and it answers whatever question you bring to it, in this country and increasingly across the world, because ownership and risk and deception are concealed the same way everywhere and exposed the same way too. Bring us a name, and we will show you the network.

Two disciplines keep all of this honest, because a capability like this is dangerous if it is sloppy. We are ruthless about proving it is the right individual and not a namesake, because a confident wrong answer is worse than no answer at all. And every result is a lead for a trained professional, not a verdict and not an allegation. The aim is never a longer list. It is a shorter, cleaner one that a professional, a journalist or a regulator can stand up.

Now the part that makes it more than an intelligence tool.

An AI worker is only ever as good as what it knows. Give it a plausible model of the world and it will act plausibly, which is to say confidently and sometimes catastrophically. Give it a resolved, evidenced picture of who owns what, who connects to whom, where risk and money actually concentrate, and whether a source can be believed, and it can act with a real understanding of the situation in front of it. Same worker, completely different creature. One is guessing well. The other is reasoning over the truth.

That is the ground beneath what we build on top of the intelligence: autonomous digital workers for the work where being wrong is expensive. Not a chatbot bolted onto a helpdesk. Workers with a persistent identity, a genuine and consistent character rather than a prompt wearing a costume, their own judgement about when to act and when to stop and bring in a person, and beneath all of it a true picture of the world they are working in rather than a confident hallucination of it.

And in regulated work, that raises the question everyone in this field is quietly hoping you will not ask. How do you know it behaved? "Trust us, it complies" is not an answer a bank, a regulator or a court will accept, and it never should be. So we hold ourselves to a harder standard, and we have a name for it: provably compliant autonomous agents.

Here is what that actually means, in plain terms. In most systems, compliance is a filter bolted on beside the agent, an outer box you hope it cannot climb out of. Ours is not built that way. The rules the agent must obey are part of the act itself, not a filter beside it, so there is no outer box to escape because the constraint travels inside the agent's own decision to act. The compliance properties that matter are mathematically proven to hold, not sampled across a test set and hoped, and proof is a far stronger claim than a passing test. And every single action the agent takes carries the evidence that it stayed within its mandate, so when the question comes, months later, from an auditor or a court, you can show exactly what happened and that it could not have happened otherwise.

Compliance you can prove, not a promise you make. That is the whole philosophy in one line, and it is the same line as the intelligence work underneath it: others give you plausible, we give you provable. A finding you can stand up, from a system you can audit, doing work you can account for.

A country that cannot see who owns what cannot enforce its own rules. An organisation that cannot prove why its systems did what they did cannot trust them with anything that matters. Both problems have exactly the same answer, and it is not a cleverer guess. It is the discipline to resolve the truth, and to show your working, all the way down.

Plausible is easy, and everyone in this industry is selling it. Provable is hard, and it is the only thing worth building.

More on the workers themselves, and what this makes possible in practice, next.

Read it, cite it, argue with it

Read it, cite it, argue with it

• The hidden property ownership story, in Private Eye and written up in full
• Kronaxis Intelligence
• Provably compliant agents, the architecture
• DYNAMICS-8, the science under the personas' character (open access)
• All four papers, open access with citable DOIs