Main finding
The research frontier has moved from producing fluent output to representing how an environment changes and acting inside it. A World Model learns the behaviour of a system so it can be rolled forward as a simulator. The approach is well demonstrated in control tasks; its behaviour on financial markets is not established.
What supports it
- Early work showed agents can be trained inside a learned model of their environment rather than the environment itself, which is what makes simulated planning possible (Ha & Schmidhuber 2018).
- A single world-model algorithm reached strong performance across many control tasks, later published in Nature (Hafner et al. 2025).
- Capability in large models tracked training compute and data volume, and the same scaling logic now governs reasoning and action quality (Brown et al. 2020).
- Tool use turns a language model into a system that takes multi-step action in the world, which is the mechanism behind autonomous agents (Yao et al. 2022).
What this does not claim
- Public evidence is strongest in games, robotics and control. Financial and organizational environments are partially observable, adversarial and non-stationary.
- A confident simulated path is not evidence that the same path will occur.
- Capability shown in a benchmark is not capability delivered in production; the gap has to be measured on our own tasks.
Planning assumptions used
- The licensed platform operates today and exposes an interface we can evaluate on our own data.
- No supplier claim is accepted from a demonstration; each one is tested against a stated criterion.
Open before signature
- Which specific capability would change a decision we make today?
- What is the evaluation set, and who signs it off?
- What has to be true for this to be a working system rather than a demonstration?
Sources
Figures marked illustrative are planning assumptions. This note separates published evidence from project assumptions and does not present a supplier quotation, a valuation or a forecast of returns.