REV JUL 30, 2026

An industrial lab for regulated decisions

World models that hold.

You have an idea, and a system that says it cannot be done. We find the way through.

Since 2008, inside banks, insurers, and health: getting improbable things through institutions that were certain they were impossible. A deadlock is almost never politics. It is that nobody in the room can see what happens next.

Start a sprint → Building regulated worlds →

“[…] I do not understand how you can even think of building an agentic system without [it] having the ability of predicting the consequences of its actions.”
Yann LeCun, on today's agents · Watch
FIG. 0: WORLDS, DISTILLED
FROM THE STATE SPACE
00

Eighteen years of finding the move

Since 2008: trading floors, property insurance, payer and provider data, claims, lending, defence. Long enough to know that the thing blocking a project is rarely the thing everyone is arguing about.

Which is the same job as building a world model. A deadlocked system is one where nobody can predict the consequences of a move, so every option looks equally reckless and nothing happens. Model the rules, run the move, and the argument stops being about opinion.

  • 2008 onwardBanks and trading floors
  • Regulated riskProperty insurance
  • HealthPayer, provider data, claims
  • AlsoLending, defence
01

Principles

Three, and none of them needs a glossary.

P-01

If you cannot explain it, you do not understand it

Feynman's rule, and the one that catches you out in a meeting. An explanation a smart outsider cannot follow is a hole in our thinking, not in theirs.

P-02

If you cannot demo it, you cannot build it

Running beats convincing. A demo is the cheapest test that an idea survives contact with the world, so it comes first, and it is allowed to be ugly.

P-03

Intersect the unconnected

The answer is usually sitting between two fields that never talk. Going and reading the other one is where the unstuck moves come from.

02

Consequences are the part nobody tests

LeCun's bar for a reliable agent is that it can predict the consequences of its actions. Almost every model in production is graded on something else: whether it got the next number roughly right, on data that looks like the data it was trained on.

In a regulated business the consequences are published in advance, with dates. That makes the bar testable. Three tests, none of which needs ground truth, a holdout set, or retraining. A model can pass an accuracy dashboard forever and fail all three.

  1. T-01 detour Does it survive a detour?

    Close a street it never trained on and see whether it still gets home.

  2. T-02 definition Does it know the ruler changed length?

    A definition moves on a published date and nothing real happens. Most models report that the world got worse.

  3. T-03 restatement Does it survive a restatement?

    Numbers you already acted on get rewritten behind you. That is normal here, and it is not in your simulator.

Run the experiments →

FIG. 1: A WHOLE WORLD FROM ONE SEED, GROWN FROM A FEW RULES. DRAG TO SPIN, SCROLL TO FLY DOWN. FULL SCREEN →
03

Work with us

Every engagement Hammer Labs takes is unique, so there is no package on this page to buy. We build a world for your problem specifically, and the only way to do that is to understand your domain deeply enough to model it.

Tell us the problem and the system it is stuck inside, and you get a scope, a price, and a date back. If it is not worth building, Hammer Labs will say so instead.

Email kartik@multiversal.ventures →

  1. Bring The problem, and what it is stuck inside

    The rulebook, the data you actually have, and the constraints everyone treats as walls. Those are the useful part.

  2. We do Learn the domain, then build its world

    No template gets reused. The model is built from your rules, your dates, and the way your system actually behaves when pushed.

  3. Get A way forward you can defend

    The move, and the reason it works, in a form you can rerun, inspect, and hand to whoever has to approve it.