Continual learning · World models

Machines that continually learn the laws of their world.

We build systems that infer the causal rules of an environment from a handful of interactions, then lock every rule they can verify in place for good.

The problem

Today's systems don't accumulate.

A modern model can absorb an ocean of data and still hold nothing it can stand behind. What it learns is rarely checked, and what does get checked is rarely kept — the next round of training quietly erases it.

So progress leaks. Capabilities appear and vanish, correct behavior decays into confident error, and there is no ledger that says which gains are real and which have already slipped away.

Scale alone does not close this gap. A larger model forgets on a larger budget. The missing piece is not more parameters but a mechanism that verifies what has been learned and refuses to let it slide back.

Our thesis

Three commitments, one system.

World models

A learned simulator of the environment you can query and intervene on — ask what happens next, or set a cause and read off its effect — not a black box that only predicts the surface.

Causal · queryable

Program synthesis

Each law is written as a discrete program — a small, inspectable rule you can read, test, and check against reality — instead of a weight buried in a tensor no one can audit.

Discrete · checkable

The ratchet

Continual learning where every verified law is caught and held. New knowledge can be added; proven knowledge cannot be lost. This is the difference — the pawl that turns learning into accumulation.

One-way · locked in

The name

Why "ratchet."

A ratchet is a toothed wheel and a pawl — a small catch that rides over each tooth as the wheel turns forward, then wedges against it the instant the wheel tries to turn back. Motion is permitted in one direction only.

That is the property we want in a learner. Turn the wheel forward with every interaction; let the pawl drop into place the moment a law is verified. Gains are earned once and held for good — no round of training can ever unwind them.

Join us

We're early. Come build it with us.

Ratchet Labs is a young research lab. If efficient causal learning, program synthesis, or continual learning without forgetting is the question that keeps you up, we want to compare notes and build together.