Research profile

World State and JEPA

Grounded world models that distinguish observations, inferred state, predictions, actions, and outcomes.

World models

Grounded world models that distinguish observations, inferred state, predictions, actions, and outcomes.

The World State research track investigates how a persistent agent system can maintain beliefs about an environment without presenting predictions as observations. Its architecture keeps external evidence, inferred state, action-conditioned futures, proposed acts, and grounded outcomes in distinct records. Compacitas defines a model interface for structured baselines, JEPA, video-world, physics, and hybrid approaches. Candidate models must be executed and evaluated before promotion, with replay and calibration evidence tied to their permitted operating scope.

The research question

Which world-model representations improve action-conditioned prediction while keeping uncertainty and observational grounding explicit?

Research workstreams

A question. A method. A test.

Workstream 01

Observation–state–prediction–action–outcome data contracts

Workstream 02

Evaluation of structured, JEPA, video, physics, and hybrid backends

Workstream 03

Replay, calibration, failure boundaries, and admission evidence

Evidence priorities

Make progress measurable.

The initiative is framed around testable questions. Its research direction should be assessed against observed evidence and the limits of each experiment.

Action-conditioned prediction and uncertainty calibration
Separation of observations, inferences, predictions, and outcomes
Replay and failure boundaries before admitting a model to use

Research goals and proposed mechanisms are not established findings. Product-linked tracks describe the research behind a capability, rather than a claim that its hypotheses have been validated.

Associated company