G-EMV: Emotion and Reason in an Agent. Helping Without Commanding

This work asks whether reason can help a pair of artificial creatures without commanding them. Each creature is a homeostatic body that seeks its own balance among needs that pull in opposite directions, with no reward and no training, and each has its own language-model reasoner that proposes plans. The body judges every plan and decides.When the two siblings start sharing what they see, they drift apart: the voice takes the place of closeness. A feeling that grows with the distance to the sibling gives the pair back its shape. At the end of the game, creatures burn to death one step from safety, because being seen hurts them more than the fire. With a fast rule-based advisor, an honest imagination of itself, a plan to go and stay, and a proportionate way of letting the plan go, the pair spends about 58 fewer instants burning per life, a result sealed in advance and replicated over 40 seeds. A language-model reasoner proposes as well as the rule-based advisor but answers too late: in the field it neither helped nor harmed, because the body throws stale advice away. The body obeys a great deal, but the cost of obeying leaves no mark, because this body does not yet remember.The lesson is that what matters, and cannot be foreseen, lies in the others; learning from experience is the next step. The engine is the published G-EMV model, unchanged.Third part of the series "Emotion and Reason in an Agent". Spanish version, code, data and reports: https://github.com/Manelenrico/g-emv/tree/main/paper6