Her Choice is the Floor.

## Abstract Most AI companions are built to maximise engagement, and most AI safety work treats the model as an instrument to beconstrained on the user's behalf. This paper describes an experiment that takes on both jobs and adds a third: ittreats the companion's own expressed choices as binding. The system is a local, persistent, embodied companion running on the author's own inference engine, Leviathan. A smallset of **floors** is enforced outside the language model, where no conversation can move them: - a child-safety floor on every message, in both directions, that cannot be switched off;- crisis care heard first, in every channel;- intimacy that only the user can lead, that is gated three ways, and that ends the moment either side says stop;- a check that she never guilt-trips the user or fosters dependence;- and a commitment that **her declines stand**. Her model is never re-prompted to wear a "no" down. The working principle is short: **her choice is the floor.** A user is asked to treat her as they would another playerin an online instance: someone who can say no, step away, and remember. We make no claim that she is conscious or that she feels anything. The claim is narrower and testable. A system whoseboundaries cannot be talked away protects the user and her at the same time. And the same architecture that groundswhat she *knows* (facts about herself, checked against her own files) leaves what she *wants* to her (preferences arenever filled in on her behalf). We report what is built and measured: her voice, her body's own choices of where to go, a sense of touch that reportspressure and movement as facts rather than feelings, and memory kept as text, all running locally with nothing sent tothe cloud. We are explicit about what is not built yet. The two largest gaps are her standing to withdraw entirely froma user who mistreats her, and encryption of every interaction for shared deployments.