Atlas
Discovery, separated
from reproduction.
A small, structured, verifiable world-model framework for physics identification research — an auditable counterpoint to large, opaque video world models.
Atlas
in one place.
Atlas is a small, structured, verifiable world-model framework for physics identification research.
It is designed as an auditable counterpoint to large, opaque video world models. Its latent space is deliberately decomposed into interpretable components, and its claims are tied to reproducible scripts, regression tests, evidence records, and documented retractions.
The reason
it was started.
A model can generate accurate predictions while still giving a false impression of what it has learned.
It may reproduce a quantity without identifying the underlying mechanism. It may encode static and dynamic information in entangled ways. It may appear to infer physics while actually exploiting coordinates, leakage, shortcuts, or engineered signals.
What stands
in the way.
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.
How the work
is approached.
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; probes what physics can be identified from observations; and preserves failed claims when results do not survive verification.
What it can
contribute.
Genesis cannot claim to construct its own operational reality unless the laboratory can distinguish discovery from reproduction.
Atlas contributes the epistemic discipline required for that distinction. It helps Genesis ask whether a relation was truly present in the observations, whether the architecture forced the representation, and whether the claimed distinction survives intervention, reproduction, and transfer.
The state
today.
Active research. The framework, its experiments, and its evidence records are maintained in the open repository.
Claims are held to reproducible scripts and regression tests, and retractions are preserved rather than removed when a result does not survive verification.
- 01Type — World-model research
- 02Status — Active research