Hammer Labs: world models
World models that hold. An industrial lab for regulated decisions.
When a system says an idea cannot be done, Hammer finds the way through. World models of regulated systems, so software is tested against the rules instead of around them.
The pitch
When a project deadlocks inside a bank, an insurer or a health system, it is almost never politics. It is that nobody in the room can predict what happens next, so every option looks equally reckless and nothing moves.
Hammer builds world models of regulated systems: model the rules, run the move, and the argument stops being about opinion.
The offer
- 01 · BRING
The problem and what it is stuck inside: the rulebook, the data, the constraints.
- 02 · MODEL
The lab learns the domain and builds the world model from your own rules and data.
- 03 · RETURN
A defensible path: the move, and the reason it works. You can rerun it, inspect it and present it. You get a scope and a date back. If it is not worth building, we say so.
Three tests, no ground truth needed
- T-01 Detour
- Does the model survive a detour? Novel scenarios it was never shown.
- T-02 Definition
- Does it know the ruler changed length? Regulatory definitions that moved under it.
- T-03 Restatement
- Does it survive a restatement? Retroactive corrections to the data it learned from.
In a regulated business the consequences are published in advance, with dates. That is what makes the bar testable. It is the standard a reliable agent should meet: it can predict the consequences of its actions.
Proof that an agent works, and stays legal
Everyone is shipping agents. Few can prove theirs work, and fewer can prove they stay legal as the rules change. That work lives in Hammer Worlds now: simulated users built from a world model rather than canned scripts, findings a team can act on, test suites the developers keep, and the same harness rerun on every release so an agent is measured against the business bar.
Research that ships
The lab publishes real engineering: KV-cache compression ported into a CUDA fork and run in production document work; business-aware loss functions that weigh the true cost of being wrong, not only accuracy. Experiments run in the open and the numbers are reproducible from the code.
What a world model does not claim
- It does not predict what a regulator will decide. It says what the published rules entail on a date. A ruling, an audit finding or an adjudication is a decision somebody else makes.
- It does not pass an agent that fails its own tests. The three tests are the bar. An agent that fails a detour, a definition change or a restatement is reported as failing, whoever built it.
- It does not stand in for your data. A model built from the public record answers from the public record. Your contracts, your claims and your outcomes are evidence only you hold, and a private world is what we build with you.