Research · World Lending Model
A world model for lending.
A learned model of how lending actually behaves, so every agent from origination to collections plans against the same picture. No action fires until the model has simulated it and the compliance gate has cleared it.
The idea
Borrowed from the frontier of AI.
A world model is the AI’s picture of its environment, detailed enough to play an action forward and see what happens. The aim is for an agent to simulate an action against the model in a sandbox, checking the likely borrower response and compliance risk, before taking it for real.
Lending is where that matters most. Its decisions are consequential and hard to reverse, so testing one before it happens is worth more here than anywhere.
The architecture and the rulebook agree
“Check an action against its consequences before you take it” is the world model’s safety property. It is also, in essence, what a regulator asks of a credit decision.
It is the same control in both worlds, which is why a world model is the right way to build lending that scales without scaling its risk.
Inside the model
Many models of the world, learned as one.
The world of lending is too large for one function to capture. The World Lending Model builds it from parts, each learned from real outcomes, each sharpening the others.
The borrower model
A living read on a borrower’s ability and willingness to repay, and how it shifts with income, behaviour and life events. It moves as their life does, so it never goes stale like a fixed credit score.
Products & cashflows
How a loan’s terms, balances and repayments behave over its life, under each path a borrower might take. It links the way a loan is built to the economics it produces.
Market & macro
The world outside the loan: interest rates, employment, sector and regional conditions. It captures how a shift out there moves both borrower behaviour and portfolio outcomes.
Policy & rules
The rules in a form the machine can check: fair-lending law, each jurisdiction’s regulations, the institution’s own risk appetite. Every proposed action is tested against them before it goes ahead.
Lifecycle dynamics
How a case moves through origination, underwriting, servicing and collections, and how one action at any stage changes what is possible at the next.
+ the agent harness
The layer that carries the model’s predictions into the real world, enforcing every compliance boundary as it acts.
The agent harness
A powerful world model, made safe enough to act.
Agents work the whole stack: origination, a validated AI underwriter, servicing and collections. Every action, including each credit decision, is scored against the model and then cleared by the compliance gate before it can fire. It carries the reasons it was approved, and its outcome feeds back so the model gets truer.
A model is only as good as its accuracy, so validation and learning are part of the harness itself. Validated before it acts, explainable in its reasoning, corrected by real outcomes: that is what makes an AI underwriter fit for regulated credit.
Validated before it acts
Every proposed action clears Krim-Nyāya, our pre-execution validator, against local law, policy and the institution’s risk limits before it can fire.
Explainable by construction
Every decision carries a step-by-step reasoning trace, recorded immutably in Krim-Ledger: cryptographic proof on demand, ready for any audit.
Corrected by outcomes
Krim-Learn, our learning loop, routes payment outcomes back into the model, so its accuracy improves the more lending it sees.
Sovereign
It reasons, acts and improves entirely inside the institution’s own walls.
End to end
One model, the whole of lending.
Origination, the AI underwriter, servicing and collections all reason against the same model, every action proven before it acts. That breaks the wall between the front and back of the book: collections data retrains origination, and a borrower’s underwriting file guides the recovery conversation later. One shared intelligence keeps every decision accountable to the regulator, and creates room to serve more borrowers safely where the risk genuinely supports it.
Origination · with the model
Each application is read against everything the model has seen, so a thin file or a new segment can be judged on behaviour, not a bureau score alone, and approved where the risk genuinely supports it.
One model, the whole lifecycle
The model that gets better the more lending it sees.
Built on Kendra, the runtime that validates every action and learns from every outcome.