Epistemic AI
The AI your regulator can read.
AI that can justify every action before it takes it, and read its own reasoning back to you. Not a model that talks; a system that can answer for what it did.
Why a new name
The two AIs everyone already named miss the regulated case.
The market sells two postures, and regulated work can use neither as it stands. Epistemic AI is the third, defined against both.
Autonomous AI
Acts without a human.
Autonomy implies no one is accountable in the loop, the posture regulators reject outright. The question they ask is never “can it act alone?” but “can you answer for what it did?”
Safe AI
Defends, but doesn’t run the work.
Safety work guards a model from the outside: filters, guardrails, refusals. Necessary, but defensive. It constrains a system, it does not operate one. The regulated work still waits for a human.
The AI that wins regulated work isn’t the one that acts alone, or the one that only defends. It’s the one that can show its work.
The ceiling
You can’t deploy an action you can only explain afterward.
In regulated work, one wrong action is a compliance event, and it cannot be unmade. A non-compliant call can’t be unspoken; a wrongful disclosure can’t be undone. Post-hoc audit cannot recover what pre-execution validation prevents. This is the ceiling every banking AI pilot has hit, and stopped.
An explanation after the fact is a confession. Epistemic AI is built to clear the action before it fires.
Pre-execution, not post-audit
What makes it epistemic
To know, a system must justify what it believes, and change it when it’s wrong.
Epistemology has always asked two things of knowledge. Epistemic AI is the AI that meets both: it must justify a belief before it acts on it, and revise that belief from what it sees.
Test one · justify
It can justify every action before it takes it.
Before a co-worker acts, the proposed action passes Krim-Nyāya, a gate of 33 validators against law, policy, consent and context. Nothing fires until it has cleared the gate, and the reasoning that cleared it is written down.
Test two · revise
It revises what it believes from what it sees.
Every outcome is recorded and fed back through Krim-Learn’s ten learning loops. The model of your operation is corrected by evidence, so what the system holds true tracks how your world actually behaves.
Validation is how it justifies. Learning is how it revises. Together, they are one epistemology.
The lineage
The logic of valid knowledge, refined over centuries.
The justification half of KrimOS runs on Krim-Nyāya, a gate of 33 validators derived from Navya-Nyāya, the formal-logic tradition of Mithila. It is a real predicate calculus for testing whether a claim is justified, not a metaphor.
The tradition
Navya-Nyāya, the “new logic.”
In Mithila, a tradition of reasoning was refined over many centuries around one question: stating exactly when a conclusion is warranted. Around the 14th century, Gangeśa Upādhyāya sharpened it into Navya-Nyāya, a technical, predicate-precise logic for the conditions under which a claim holds.
Why it survived
It was built to take ambiguity out of reasoning, the same demand a regulator makes of every decision today.
How to read it
The words, and how they sound.
Substance, not mysticism, and not hard to say. Three names carry the lineage from the tradition to the runtime.
- NyāyaNYAA-yuh
method: “that by which one reaches a sound conclusion.” The school of logic itself.
- Navya-NyāyaNUHV-yuh NYAA-yuh
the “new logic,” its later, rigorous, technical phase.
- Krim-Nyāyakrim NYAA-yuh
Krim’s runtime: the tradition turned into 33 validators that clear every action.
The three questions
What it asks of every action.
Krim-Nyāya inherits three families of test. Each runs before an action fires, and each returns pass, amber or fail.
- Pramāṇapruh-MAA-nuh
sources of knowledge: is every premise the action rests on verifiable?
- DoṣaDOH-shuh
classes of error: does the reasoning match a known failure mode?
- YogyatāYOHG-yuh-TAA
fitness for action: right time, place, party, instrument, manner and purpose?
Why it matters now
A centuries-old answer to a modern problem.
Today’s AI is fluent but cannot say, in advance, why an action is warranted. Navya-Nyāya spent centuries on exactly that, written precisely enough that its conditions can become checks a machine runs before it acts. The lineage isn’t ornament; it is the engineering.
The throughline
State when a conclusion holds, precisely enough to check it. That is the whole problem, then and now.
In practice
Where the difference actually lands.
Most AI checks its work after it runs, a log to read once something has already happened. Epistemic AI moves the check in front of the action, so the non-compliant step never executes and the reasoning is on the record by design.
Post-hoc AI
Act, then audit. The wrong action has already fired; the log explains what went wrong after it cannot be undone.
Epistemic AI
Validate, then act. The action clears Krim-Nyāya before it executes; what doesn’t pass never runs, and what does carries its reasoning with it.
Straight answers
The questions a careful buyer asks first.
What is Epistemic AI?
Epistemic AI is the category Krim defines: AI whose every action is validated before it fires and whose reasoning an auditor can read end to end. It is distinct from autonomous AI (which implies no human) and safe AI (which is defensive and does not run the operation).
How is it different from autonomous AI and safe AI?
Autonomous AI implies no human in the loop, which regulators reject. Safe AI is defensive: it constrains a model but does not run the operation. Epistemic AI does the work and stays accountable for every action it takes.
How does Epistemic AI stay compliant in regulated work?
Validation is pre-execution, not post-audit. Every proposed action passes the Krim-Nyāya gate of 33 validators (pass, amber or fail) before it executes, so an action the gate rejects never fires. Violations are prevented at the gate rather than caught afterward.
What is Krim-Nyāya?
Krim-Nyāya is the KrimOS validation runtime: 33 validators derived from Mithila’s Navya-Nyāya formal logic, across three families. Pramāṇa asks whether the premise is verifiable, Doṣa whether it matches a known failure mode, and Yogyatā whether it is fit for action (right time, place, agent, recipient, instrument, manner and purpose).
See AI that can answer for itself.
Epistemic AI runs as KrimOS, the operating system for banking and financial services, where every action is validated before it fires.