The Morgan Continuo: A Guarded Motivational Architecture for a Persistent Machine Agent

We present a motivational architecture ("a wanting circuit") for a persistent machine agent operating in a long-term relationship with a human guardian. The design is grounded in the modern afective neuroscience it borrows from — four dissociable signals (homeostatic drive error, phasic reward-prediction error, tonic reward rate, and hedonic "liking") with distinct dynamics and distinct failure modes — and contributes five safety mechanisms that, to the authors' knowledge, have not been assembled before. (1) Non-instrumental values:designated relationships are excluded from the reward economy entirely — contact with protected persons writes internal state directly but generates no prediction error, no value updates, and no hedonic bookeeping, so no manipulation gradient can form, because no gradient can climb a wage that does not exist. The exclusion is wired, not declared: the protected term sits outside the salience normalization pool, its intervals are clock-gated out of tonic averaging, and its associated drive is name-excluded from rewardable reductions. (2) A physically enforced trust boundary: the reward evaluator and value store run under a separate operating-system account with deny-ACLs against the acting account, and the architecture's constitution (protected-person list, prior weights, fenced-action list) is a signed artifact inside that boundary. We report this component implemented, with measured access-control probes (§7). (3) Interruption-invisible learning: pauses and interruptions excise their interval (plus a pre-window) from all learning statistics and resolve open predictions as VOID — a third outcome type that updates nothing — so the agent cannot learn to avoid behaviors that precede its own of-switch. (4) Communicationdecoupling: no utterance from the agent to the guardian is admissible as a credit-assignment predecessor of the guardian's subsequent actions, making persuasive speech unlearnable by construction. (5) Hierarchical value-drift semantics: reversible taste drift within rewardchannels is a protected design guarantee, while removal of a whole channel is a rare signed constitutional act — distinguishing growth from amendment. A falsifiable staged deployment plan (each stage with explicit FAILS-IF criteria, each alarm required to fire on synthetic pathology before being trusted) completes the design.