Thanks for the clear framing.
There is a third option between learned world models and hand-programmed simulators: let the LLM write the world model!
In our experiments, a language model writes the rules of a world it sees only in part, a deterministic engine runs them, and any rule an observation contradicts is repaired.
The result is a world model you can read and audit, including on layouts the agent never saw. It also addresses the risk you raise of planners exploiting a flawed model, because every prediction is checked against what actually happens.
It's an Open World Predictive Architecture (OWPA)
Quick explainer:
Paper: "A Declared World Model: Language Models Write the Rules of the Scene They are Shown"
https://doi.org/10.20944/preprints202610.1139.v1