What Epistemic AI means.

We chose a gentle word for it. When an AI system asserts something false with total confidence, we call it a hallucination — a trick of perception, almost endearing. In a chat window it is a wrong answer you catch and correct. Move the same behaviour to the desk where an action carries legal weight, and the gentle word falls away. A hallucinated right-party contact is a call placed to the wrong person. A hallucinated adverse-action reason is a notice that breaks the law. The model didn’t know it wasn’t allowed to do that. It was never built to know.

That is the gap this essay is about, and it is not a gap in capability. The models are extraordinary. It is a gap in kind. To see it, look at the two kinds of AI the industry has actually built.

The two AIs we have

Generative AI is trained to be plausible. Given everything it has read, it produces the most likely next word, the most convincing image, the answer that best fits the pattern. Plausibility is its whole objective, and it is spectacularly good at it. But plausible and correct are not the same thing, and plausible and permitted are not even close.

Agentic AI is trained to act. Give it tools and a goal and it will take steps in the world: call an API, move a record, send a message. It is generative AI with hands. And that is exactly the problem, because the thing we bolted hands onto was optimised to be convincing, not to be allowed. An agent that is plausible and can act, with nothing between the intent and the execution, is precisely the system a regulated institution cannot ship. Not because it is not clever enough. Because no one can prove, before it acts, that the action was permitted.

The missing third

Epistemology is the branch of philosophy that asks what we are justified in believing: what we have grounds to hold, and why. Epistemic AI is the category built around that question. It is the category Krim defines: AI whose every action is validated before it fires, and whose reasoning an auditor can read end to end. Where a generative model asks “what is the likely next token,” an epistemic system asks a different question of every action it is about to take: is this justified — are its premises verifiable, is its reasoning free of known error, is it fit to do here, now, to this person?

This is a different place to put the intelligence: in the judging of what may be done with what a model generates, not in the generating alone. A gate between the output and the act. The model can still be a black box. The decision to act on its output does not have to be — and what the system learns from each recorded outcome sharpens the judgement it brings to the next.

Where the method comes from

A category needs a method, and Krim’s is older than computing. The Nyāya tradition of Mithila is two millennia of precise reasoning about what follows from what, and what a claim must satisfy before it can be relied upon; around the fourteenth century it was sharpened into Navya-Nyāya, its technical, predicate-precise form. KrimOS turns that inheritance into a working gate. Its validation runtime, Krim-Nyāya, runs every proposed action through 33 validators, grouped into the tradition’s three families of test:

Pramāṇa asks whether the premises hold — is every fact the action rests on actually verifiable? Doṣa asks whether the reasoning matches a known failure mode — the errors we have seen before and named. Yogyatā asks whether the action is fit to take at all: the right time, place, party, instrument, manner and purpose. Each validator returns pass, amber, or fail. Nothing executes on an amber or a fail. What passes carries the record of why it passed.

A word of honesty, because the category deserves it. These validators are a grammar, not a guarantee — a disciplined way to ask, of every action, whether it is justified. They cover the harms they are written to cover, and the work of a validation runtime is never finished. The claim is not that the gate makes an action safe. The claim is narrower and more useful: that the action, and the reasoning that cleared it, are on the record before anything happens — where a supervisor, a board, or an examiner can read them.

Why banking asked for this first

Every industry will eventually need epistemic AI, but banking needs it now, because banking already lives by an epistemic standard. A regulator does not ask whether your model is accurate. It asks whether you can show the decision was allowed — this customer, this rule, this basis — before it was made, and on a record someone else can check. That is not a question about capability. It is a question about justification. It is epistemology with a statutory deadline.

Read the rulebooks in that light and they stop looking like separate regimes. The US CFPB holds that a model being too complex is no defence for failing to give a borrower specific reasons for a denial. The Federal Reserve’s SR 11-7 expects a model validated before you rely on it. The EU AI Act puts credit scoring in its high-risk tier and demands human oversight while there is still a decision to govern. The UK’s Consumer Duty asks for evidence of good outcomes, measured and shown. Different words, one demand: show that the action was justified, before it acted, in a form a third party can read. Pre-execution validation is not the letter of any one of these laws; it is the architecture that most cleanly answers all of them at once.

Generative AI made machines articulate. Agentic AI is making them act. The work that runs the world — lending first, and everything shaped like it — needs the third thing: machines that can prove they are allowed to act, and show their reasoning for it. That is the category KrimOS is built on, and it is the one an examiner, a court, and a board were always going to require. Not the AI that sounds right. The AI your regulator can read.