The model suggests, code applies

A small simulation of a village, with a local model deciding what the villagers want to do. The rule is written at the top of the project: never let the LLM mutate world state directly.

The model returns an intent as JSON. A validator reads it. Python applies it, or does not.

When the model suggests something impossible, the validator refuses it and the world does not change. When the model is switched off, the village carries on without it.

It is a toy. It is also the same boundary as an agent in a cloud account, drawn where the stakes are a villager’s lunch. That is the cheapest place to learn where the line goes.