G-EMV: Emotion and Reason in an Agent. Curiosity, Trust and Commitment

This work observes and tests an artificial creature made of two parts: a homeostatic body that feels what it lacks in three dimensions (its body, its resources and its bonds), and a language-model advisor beside it. The question is when thinking is of use to a body like this. We give the world time, with games twice as long, and build three pieces. A wide door lets the advisor propose a whole plan with its emotional shape, and the body judges it with its own table; this door tells good advice from random advice. Trust is earned and lost with results: an advisor that talks at random soon stops moving the body. Curiosity is installed as one more need of the body: as a trip with a destination and commitment, the body discovers almost six times more world per instant while it travels, without getting closer to danger. In a hundred games, the advisor's plans gave the body no advantage: almost all the improvement in the creature's future goes through the others, and what they will do the body cannot foresee. What remains is a division of roles that makes reason useful: the body says what it needs, reason can say where to find it, and the body checks it, commits to the plan and only lets go of it if something important happens. No training, no reward; the engine is the published G-EMV model, unchanged. Fifth work in a line of research that began with the G-EMV engine (10.5281/zenodo.21026795); it follows directly from "When Thinking Is Needed" (10.5281/zenodo.22844274). Spanish version, code, reports and data: https://github.com/Manelenrico/g-emv/tree/main/paper5