MindForm — Feedback Loop: Learning What the Consequences of One's Own Acts Mean, and Whose Answer Counts

Every node documented so far in this series changes a character in response to what happens to it. None learns from what happens in answer to what it does. Behavior closes the act–consequence loop only coarsely: the next message is read as the reception of the previous reply and trains two global sensitivities, so a character can become bolder or more guarded in general but cannot hold the knowledge that its parents disapprove when it speaks up, cannot tell a parent’s “no” from a stranger’s, and cannot let a fifth identical rejection mean anything but a fifth push in the same direction. This paper documents the Feedback Learning module, which supplies that knowledge. It is a substrate and not a pushing node: a parent’s “no” is already an experience that forms every layer, so a module that pushed traits, values or moral foundations again would count one event twice. It stores a model of other people — per voice and act, a running expectation of the response, updated by prediction error with recency-weighted evidence, and per voice, learned importance, trust and warmth — and acts only by modulating channels that already exist. It scales how hard a response lands, with what is expected landing softer and a violated expectation harder; gates how much of a verdict is internalized into values and moral foundations by the voice’s credibility, after Kelman; lets an act awaiting its answer read the way it is expected to go; and promotes confident expectations into provenance-tagged beliefs along a ladder from “my parents disapprove” through “I should not” to “it is wrong to,” the last reachable only through a voice the character trusts. Importance and trust are held apart on purpose, so a character can depend on a voice it does not believe, and we show the module doing exactly that. We exercise the module in isolation and report four loops: one damped under a predictable voice and live under an unpredictable one, one feed-forward duplication of a single lesson across two routes into perception, one in which agreement between voices raises their trust whether or not they are right, and one in which a bond deepens under uniform rejection. We also record a fifth first-class quantity the architecture computes and