Inside the
world model.
Elemental is an independent open-source lab for world model interpretability. We reverse-engineer how predictive models like Meta's I-JEPA build internal representations of the world, and release everything in the open.
The lab, in motion.
What WorldModelLens makes visible.
Observabilitytooling
An open-source layer to analyze, debug, and understand world models. Built for predictive architectures like Meta's I-JEPA, it makes internal representations observable as they form.
Replay &faithfulness
Step back through a world model's internal trajectory and inspect what it forms, separating what the model appears to attend to from what genuinely drives its predictions.
Propagationtracking
Trace how small internal errors compound across a world model's forward pass. Reproducible faithfulness diagnostics, shared as open notebooks the whole field can run.
Three projects.
One goal: make world models legible.
WorldModelLens
An observability layer for analyzing and debugging world models: replay, representation inspection, and error tracking, starting with Meta's I-JEPA.
AAF
A framework for testing whether a world model's attention truly reflects what drives its decisions. Rigorous, reproducible diagnostics.
Eval Suite
Harnesses and reproducible notebooks that standardise how interpretability experiments are measured and shared.
A small, growing group.
Curiosity over credentials.
Full team ↗
We believe in altruism, investing in early-stage researchers through mentorship, hands-on collaboration, and open work.
Why this work is urgent.
Detect deception
Inspect what a world model actually computes before deployment, rather than trusting behaviour that can hide misaligned internals.
Auditable trust
Provide auditable evidence of how decisions are made, moving beyond black-box testing toward verifiable understanding.
Inform governance
Supply the technical tools for third-party audits and regulatory oversight, bridging interpretability and policy.
Advance science
Contribute fundamental knowledge about how learned representations form and evolve inside world models.
Come observe
systems with us.
Researcher
Collaborate on world model interpretability. Contribute to open problems and co-author work.
Reach out →Engineer
Contribute to WorldModelLens and AAF. Build the tooling that makes world models legible.
Open the repo →Policy
Engage with the technical foundations of AI governance. Bridge interpretability and oversight.
Start talking →