Constrained architecture in small, controlled environments. Immutable, slow, dynamic, and controllable latents are kept apart. Actions may enter the representation only where the experiment allows it.
Scoped research · Scoped research
Discovery, separatedfrom reproduction.
A small, auditable environment for studying how a system can uncover rules in a world with constraints.

The operating question
Did the model discover a physical relation, or only reproduce a quantity it was already given?
Why the question matters
Define the claim before the result.
World-model research often evaluates prediction quality without establishing identifiability.
That creates the risk that a model predicts future states without recovering the physical variables claimed, that latent spaces are described as structured without architectural enforcement, or that experimental leakage makes a result look stronger than it is.
The method
Make the experiment auditable.
Atlas uses constrained architecture, small environments, controlled physics, reproducible experiments and explicit evidence ledgers.
Its approach separates immutable, slowly changing, dynamic and controllable latent components; restricts where actions may enter the representation; compares random and information-seeking data collection; and preserves failed claims when results do not survive verification.
Claims are tied to reproducible scripts, regression tests, and an evidence ledger. Failed claims stay in the record. Toy-environment evidence is audited; causal ablations and MuJoCo transfer remain under validation.
Current evidence
Scoped research
Active research. The toy-environment evidence core is audited; causal feature ablations and MuJoCo transfer remain under validation.
Claims are held to reproducible scripts and regression tests, and retractions are preserved rather than removed when a result does not survive verification.
04 / Limits
What is not claimed.
Results are published only within their verified experimental scope. Failed claims stay in the record.
Continue