What KXCO Is, and Why the Hard Problem Was Never Intelligence

The prevailing enthusiasm assumes the hard problem is intelligence. It is not. The hard problem is that we are asking the institutions which actually carry consequence, banks, hospitals, courts, governments, to permit software to act on their behalf while giving them no means of establishing who decided, on what basis, or whether the record will still be legible in a decade.

What we have built is not a model and it is certainly not a chatbot. It is the substrate on which an institution can allow a machine to propose, require that a named person accept, and produce an artefact that survives adversarial scrutiny, a hostile counterparty, a regulator, and incidentally the arrival of quantum computation.

The question was never whether the machine is clever. The question is whether you can defend what it did.

Forty years of the same argument

I have spent forty years around markets, and in that time I have watched four or five waves of software arrive at institutions carrying the same promise and meeting the same reception. The promise is always speed. The reception is always a question about the record.

When electronic execution arrived, the argument that won was not that a machine could route an order faster than a man on a telephone. Everybody already knew that. The argument that won was the timestamp. Once the tape could be reconstructed, once you could put a sequence in front of a regulator and show the order, the route, the fill and the clock, the machine became acceptable inside a regulated firm. Speed was the selling point. The record was the permission.

Nothing about that has changed except the size of the claim being made. This wave is sold almost entirely on capability, and the demonstrations are genuinely impressive. I am not sneering at them. I use these systems every working day and they have changed how much work one person can carry, which is a point I will come back to, because it is the most important thing in this article. But a demonstration is not a deployment, and the distance between the two is not measured in intelligence. It is measured in whether the institution can answer for the outcome.

Nobody at a bank ever lost their licence because a model was wrong. Models are wrong constantly. That is what a confidence interval is for. People lose their jobs, and firms lose their permissions, when nobody can explain afterwards how a decision came to be made. That is the failure mode. It has been the failure mode my entire career, and it is the one this industry is currently building into its products at speed.

An agent that does the work of an analyst is worth nothing to a bank that cannot say who authorised the trade.

Four questions, asked before anything is allowed to act

Any institution, before it permits anything to act on its behalf, asks four questions. Written out plainly they look almost too simple to be the obstacle.

1. Who decided? A named person, holding a role, who can be asked about it under oath.

2. On what basis? The inputs that were in front of that person at the moment they decided.

3. Under what authority? The bounded permission they held, granted by somebody who had it to give.

4. Will the answer still read in ten years? Without the vendor, the original software, or anyone who was there.

These are not questions about artificial intelligence. They predate the computer entirely. A credit committee minute from 1954 answers all four without effort: the signature at the bottom, the memorandum attached to it, the delegated authority recorded in the minute book, and the whole thing sitting in an archive in a legible hand. Paper is an extremely good accountability technology. That is why it has survived every attempt to kill it.

An AI agent, as most of this market currently ships one, answers none of the four.

It cannot answer the first, because "the model decided" is not an answer. A model is not a legal person, holds no authority and cannot be examined. Naming the vendor does not rescue it, since the vendor was not in the room.

It struggles with the second, because a chat transcript is not a basis. It is a conversation. A basis is a set of named inputs that somebody else can fetch again and check against what was claimed about them. Most of what passes for provenance in this market is a screenshot of a chat.

It fails the third almost universally, and this is the one that puts firms in front of a tribunal. Authority is not access. Software running with an operator's full credentials does not hold authority, it holds keys. The difference appears the first time something goes wrong and the firm has to explain what the software was permitted to do, as opposed to what it happened to be capable of doing.

The fourth is barely considered at all. Ten years is not a long time in litigation. It is an ordinary gap between an action and the argument about that action. Inside ten years the model is several generations retired, the format has drifted, the integration has been rewritten twice, and some of the cryptographic assumptions in use today are on a published timetable to become arguable.

Answer all four and the machine is welcome inside the building. Miss one and it lives in the sandbox forever, however clever it is. Everything KXCO does sits on that observation.

Accountability has the shape of a graph

When I have to explain this to a board, I stop writing sentences and draw a graph, because the sentences hide the thing that matters.

Rows and columns are very good at telling you what happened. They are poor at telling you how something came to happen, because that answer is not held in any single row. It lives in the path between rows. Ask a table who decided something and the honest answer it can give is a user id in a column, which records the last hand that touched the data rather than the person who took the consequence. Those are often different people. When they are the same person, you still cannot prove it from the column.

We reach for Neo4j when we need to reason about this, and when other people need to see it. Not because graph databases are fashionable, but because the question an institution asks is literally shaped like a traversal. Who decided is one hop. On what basis is a longer walk. Under what authority is a hop in a different direction entirely. Accountability, once you strip the vocabulary off it, is reachability. Can I get from the artefact in my hand back to a person, and can I write down every step I took getting there?

Figure 1. The record an institution can defend. The machine proposes, a policy bounds what it may propose, a named person accepts, and only an accepted decision produces an artefact.

Read it left to right. The agent proposes, which is the part the market finds exciting and which is, commercially, the cheapest thing in the picture. The policy sits underneath the proposal and bounds it, so the question of authority is answered by an edge rather than by an assurance. The arrow that actually matters points down from a circle with a person's name in it, and it is labelled ACCEPTED. Nothing becomes an artefact without it.

Once the record has that shape, the four questions stop being a governance exercise and become a query. Here is the first one, written in Cypher, which is the language Neo4j uses and the one I find clearest for showing people what the question actually is.

MATCH (a:Artefact {id: $id})<-[:PRODUCED]-(d:Decision)
MATCH (d)<-[:ACCEPTED]-(person:Person)
MATCH (d)<-[:BECAME]-(prop:Proposal)<-[:PROPOSED]-(agent:Agent)
MATCH (prop)<-[:PERMITTED]-(pol:Policy)
RETURN person.name, person.role, pol.ref, agent.id, d.acceptedAt

Who decided, in what role, under which authority, on which proposal, at what time.

That is not our schema, and I am not publishing our schema. It is the shape of the question, written in a language where the shape is visible to somebody who does not write software. If your architecture cannot answer a query of that shape about something it did last Tuesday, you do not have accountability. You have logging. The two are frequently confused by people selling the first and shipping the second.

The graph does something else that I have come to value more than the query. It makes absence loud. A record with a hole in it is a record that says something, and in a table a hole is a null, which is silent and looks like every other null. In a graph, if there is no ACCEPTED edge, the path does not complete. The artefact cannot be reached from a person. Nobody has to remember to check, because the question itself fails to return a row, and a query that returns nothing when it should return a name is the most informative thing in an audit.

This is why I keep drawing these pictures for boards rather than talking about architecture. A director does not need to know how anything works. A director needs to know which circle has their name in it and what happens if the arrow is missing.

Two layers, and a gate between them

KXCO is two layers with a gate between them.

The intelligence layer reads, models, assembles and proposes. It is the part that does the work that used to take a team a fortnight, and it is very good, and it is never the last word about anything.

The accountability layer records what was proposed, who accepted it, under what authority, and produces the artefact that leaves the building. It is deliberately a separate thing with a separate job.

I am not going to explain how either layer works. That is not coyness, it is commercial sense. The mechanism is the business, our competitors read Live Trading News as closely as our clients do, and the people who need the detail get it under an agreement with their name on it. What I will give you is the shape, and then the public artefacts, so you can satisfy yourself that the shape is real rather than an illustration.

Figure 2. The gate is the design. Everything else is engineering. There is no path from the intelligence layer to the outside world that does not pass through a named person.

Notice what the drawing does not contain. There is no route from the top band to the bottom band that skips the middle. That constraint is the product. Everything else is engineering, and engineering can be bought, hired or copied.

The separation matters for a reason that took me a while to say properly. If the same system that proposes the work also certifies the work, you have rebuilt the exact conflict of interest that the record exists to remove. We record. We do not certify, and we are not the investigator. A vendor who marks their own homework has produced a marketing document with a hash on it.

The same logic runs one step further and lands on the rule I care most about. The record has to be checkable without us. If verifying what happened requires our servers to be running, our company to be solvent and our support desk to answer the telephone, then we have not sold a record. We have sold a subscription to the truth, and those expire. A counterparty holding the artefact and the public material should be able to check it on a laptop with the network cable out, ten years from now, with nobody from KXCO present and nobody from KXCO able to interfere.

The acceptance is the product

Every conversation I have about this reaches the same moment, usually about twenty minutes in. Somebody asks how we get the human out of the loop, because the human is the slow part.

I understand the instinct. I have run businesses where the approval queue was the bottleneck and the person holding it was on a flight. But the human is not in the loop for speed. The human is in the loop because a machine cannot be held to account, and an artefact that nobody accepted is a draft. Take the person out and you have not automated the decision. You have deleted it, and what comes out of the other end is an output, which is a different and much cheaper object.

This is worth being exact about, because the distinction is where the money is. An output is a thing a system produced. An artefact is a thing an institution stands behind. A valuation, a diagnosis, a filing, a term sheet, a discharge summary, a research note: these are all artefacts, and what makes them artefacts is not the quality of the analysis inside them. It is that a named party has accepted the consequence of being wrong. Quality is necessary. Acceptance is what makes it a thing you can sell, file, defend or sue over.

What changes under our design is not whether the person is involved. It is what they are doing. They are not typing. They are not reformatting, chasing attachments or rebuilding a model somebody else built last quarter. They are accepting, and one person can accept a very large amount of work in a day when the work arrives properly assembled, with its basis attached, its authority already checked and the exceptions flagged rather than buried.

Our own company runs exactly this way, and I would not sell something I do not use. Most of the work at KXCO is done by people. The software prepares, the people decide. Where the code cannot resolve something, we do not go looking for a cleverer rule. We put a clear label on it and a person signs. A gap in the software is usually not a gap at all. It is the point where somebody takes over, and a system that hides that point is lying about itself.

Take the person out and you have not automated the decision. You have deleted it.

There is a commercial argument here as well as a governance one, and it is the argument I would make to anybody building in this space. The firms chasing full autonomy are building toward a market that does not exist at the price they are imagining. Not because the technology will fail to arrive, but because removing the accountable party removes the only thing the institutional buyer was purchasing. You cannot insure it, you cannot file it, and the first time it goes wrong the firm discovers it has bought a liability with a monthly fee attached.

The question that arrives ten years late

Everything in this design is aimed at a day that has not happened yet.

The regulator opens a file on a practice that was normal when it was carried out. A counterparty wants out of a deal and their lawyers start at the weakest document. An acquirer's diligence team asks how a number in your management accounts came to be that number. A journalist has a document and wants to know who signed it. None of these people are hostile in a personal sense. They are simply doing the job of testing whether what you said holds.

By the time they arrive, the model that did the work is three generations retired. The engineer who built the integration is at a different firm. The person who accepted the decision has probably left, and may or may not remember. Your vendor might have been acquired twice. The artefact has to answer for itself, with nobody in the room who was there, and that is the standard the whole thing has to be built to.

Drawn as a graph, the question they arrive with is a walk.

Figure 3. Six hops from the document in their hand to the person who is answerable. No step in this walk requires the vendor to be alive.

That is the test. Not whether the system performed well on the day, but whether the walk still completes when everyone who was present has gone. A record that only resolves while the original software is running is not a record. It is a rented memory.

The cryptography carries a date on it too, and this part gets waved away far too easily. The mathematics protecting most records made today has a published expiry, and the relevant question for an institution is not when the machines arrive. It is whether the protection outlives the period in which the record still matters. A twenty five year mortgage file, a medical record, a trust deed, a corporate register: these outlive the assumptions they were signed under unless somebody planned for it. We took that seriously early and we did the work in public, which I will come to at the end.

Machines are an amplifier, and the amplifier needs somebody to amplify

I do not believe the future is artificial intelligence. I believe the future is people with artificial intelligence, and the difference is structural rather than sentimental.

Look at what happened to my own trade. The screen did not replace the trader. It destroyed one kind of trader and made another kind ten times more valuable. The man who was paid for knowing where the price was lost everything, because the screen knew it better and faster and for nothing. The man who was paid for judgement about the price was handed an amplifier, and his best year came after the thing that was supposed to end him arrived on his desk.

Every tool that has ever mattered has done that. It takes the mechanical part of a job, which was never the part anybody valued, and drives its cost to nothing. What remains is the part a person is answerable for, and that part gets more valuable, not less, because there is now far more of it moving through the same pair of hands.

The economics of this wave are not complicated when you put them that way. Machines have made proposals cheap. Drafting, summarising, modelling, reconciling, first-pass analysis: all of it is collapsing toward free, and it is going to keep collapsing whatever anybody thinks about it. When proposals are cheap, every bit of value in the process moves to acceptance, because acceptance is the scarce thing and the only thing anybody is paying for at the end. So build the architecture where the value went. That is the whole strategy, and it explains every design decision in this article.

Figure 4. The third panel is the business. Only it produces an artefact with a name attached, at the speed of the second.

Now the uncomfortable half, and I want to be plain about it because I employ people and I am not going to pretend this is painless. One person with a properly constructed amplifier does work that used to occupy a department. That is true today and it will be more true next year. Anybody telling a workforce otherwise is being kind rather than accurate.

But the department's work was never the typing. It was the reading, the weighing, the knowing which counterparty exaggerates, the instinct that a number is wrong before you can say why. The amplifier is very good at the first draft and it is useless at the consequence. It does not know what your firm can survive being wrong about. It has no view on whether this client is worth the risk, no memory of the last time somebody tried this structure, and no skin in the outcome at all. Those are the things that were always the job, and they are now the whole job.

So the correct posture toward all of this is not defensive and it is not evangelical. Take the amplifier. Point it at the mechanical half. Keep your name on the other half and make sure the system records that your name is on it, because that signature is the thing that cannot be commoditised and the thing the machine will never be able to supply.

Propose, accept, record

Strip the theory out and the working pattern is three words long, and it is the same three words whatever the institution does for a living.

A deal room. The documents arrive in the state documents always arrive in, which is to say incomplete and in six formats. The machine reads them, builds the picture, flags what contradicts what, and drafts the summary a partner would otherwise spend two days assembling. The partner reads the summary, changes two things, and accepts it. The record holds what was proposed, what the partner changed, what they accepted and under whose authority the room was opened in the first place. The counterparty gets an artefact rather than a folder.

A research note. Our ontology engine, which we call the Round Table, assembles the position from the filings and the primary sources rather than from what somebody wrote about the filings. Every figure carries a link back to the document it came from. When a figure cannot be sourced, the engine does not round it, soften it or hedge the sentence around it. It refuses to publish the claim. That refusal is the feature. Anybody can produce a confident paragraph. Producing a paragraph that will not go out until the number resolves is the harder engineering problem and the more valuable one.

A clinical record. Same pattern, higher stakes, and the ordering matters more. The machine proposes, the clinician accepts, and what the clinician accepted is what enters the record. Not what the machine suggested, not a blended summary of the two. The accepted version, with the proposal preserved alongside it so the difference between what was suggested and what was signed is visible rather than lost.

A signature. Somebody has to sign something, which sounds trivial and is where most institutional processes quietly break. Who was invited, who opened it, what version they saw, what they were permitted to see, what they signed, and whether that still checks out years later without the platform's cooperation.

Different industries, one pattern. The machine goes first because it is fast and cheap. A person goes second because a person can be asked about it afterwards. The record goes third and permanently, because the first two are worthless to an institution without it.

We published the plumbing on purpose

I said I would not explain how the layers work. I do not need to, and this is the part I would want to see if I were reading somebody else's version of this article.

The foundational pieces are published, publicly, under our own name. They are installable, versioned and dated. You can pull them, read what they depend on, see when each version went out, and form your own view without a sales call, a demonstration environment or a word from me. That is a stronger claim than any architecture diagram I could draw, because it does not ask you to believe anything.

The packages, published under the KXCO account on npm, and worth counting yourself: npmjs.com/settings/kxco/packages

The same work seen through an independent supply chain scanner that has no relationship with us and did not ask our permission: socket.dev/npm/package/kxco-post-quantum

That second link is the one I would open first if I were you. A vendor's own page tells you what the vendor wants you to know. A third party reading the package tells you what is actually in it, what it pulls in, and what that dependency does when nobody is watching. We put the work where that kind of inspection happens because a record layer that cannot survive being inspected has no business being sold to a bank.

What I will not do in a public article is walk through the intelligence layer or the internals of how the accountability layer produces and protects what it produces. That is the part clients pay for and competitors would like for free. The shape is above. The evidence that the shape is real is at those two links. The detail is available under an agreement.

Where I think the consensus is mistaken

Autonomy is being demonstrated to the wrong buyer. An agent that books the trade by itself impresses a venture audience and terrifies a chief risk officer, and only one of those two signs enterprise contracts. The demonstration that wins institutional business is not the one where the machine acts alone. It is the one where somebody opens the record afterwards and every question has an answer waiting.

Logs are being mistaken for evidence. Almost every platform in this market will tell you it keeps an audit trail. Ask the follow-up question: who can change it, who can read it without you, and what happens to it if the vendor is acquired. A log the vendor controls is a log the vendor can be compelled to produce, edit or lose. Evidence is what survives the vendor having an interest in the outcome.

Governance keeps arriving last, as a layer bolted on top after the capability is built, usually in response to a procurement questionnaire. It does not work in that order. If the record is not in the design from the first day, what you get is a reporting feature that describes the system rather than a record that constrains it, and the difference becomes obvious at precisely the worst moment.

Nobody is pricing the expiry on their cryptography. Records being created this year are expected to hold for decades, using protection with a published timetable against it. This is not a prediction about when large quantum machines arrive. It is arithmetic about how long your records have to last, and the two numbers overlap for anything in mortgages, medicine, trusts or corporate registers.

And the whole conversation confuses the model with the system. The model is a component, it is rented, it improves every few months, and it will be replaced. The system is what surrounds it: who may ask it, what it may touch, who accepts what it says, and what remains afterwards. Anybody whose product is a model is renting their differentiator from somebody else. The surrounding system is the part that belongs to you.

Institutions adopt at the speed of their ability to defend

They never adopt at the speed of the technology. Every forecast that assumes otherwise has been wrong for forty years, and the people making those forecasts keep mistaking a pilot for adoption. The bank does not move when the capability arrives. It moves when its own lawyers stop objecting, and the lawyers stop objecting when they can see how the file will look when somebody hostile reads it.

So the useful work in this cycle is not making the machine cleverer. Other people are doing that, they are very well funded, and they are going to succeed. The useful work is making defensibility cheap, because defensibility is the binding constraint on every serious deployment in every regulated industry on earth, and it is the one nobody wants to build because it does not demonstrate well in a fifteen minute meeting.

What KXCO sells is software. We license it to institutions, they run it, and they operate it under their own permissions with their own people and their own chosen infrastructure. We do not operate their business, we do not hold their assets, and we do not stand between them and their regulator. The guarantees are properties of the software rather than promises from us, which is exactly why they continue to hold when we are not in the room. That is not modesty about our position. It is the point of the design.

The machines are going to keep getting better. That has been the safest prediction in this business for a decade and it remains safe. What will not change is that somebody has to be answerable, and that the institutions carrying real consequence will only move as fast as they can answer for. Build for that and the cleverness takes care of itself.

The question was never whether the machine is clever. It is whether you can defend what it did.

Shayne Heffernan, Ph.D., is the founder of Live Trading News, the KnightsBridge Group, Knightsbridge Law and the KXCO.ai ecosystem spanning post-quantum cryptography, identity, attestation and enterprise ontology.

This article is commentary and general information. It is not investment advice, legal advice or a recommendation to buy or sell any security.