1. The Problem: Emotion You Cannot Audit

A therapeutic agent that cannot perceive emotion is clinically inert. A patient describing escalating hopelessness, blunted affect, or a sudden swing toward agitation is communicating signal that the agent must register, retain across sessions, and let shape its own responsiveness. But the dominant way to give a language model emotional behavior, prompt conditioning and emergent affect from the weights, produces conduct that is non-deterministic, undocumented, and impossible to reconstruct after the fact.

Every regulatory regime that governs mental health software treats that opacity as disqualifying. FDA Software as a Medical Device review, the IEC 62304 software lifecycle standard, and ISO 14971 risk management all assume that a clinician or auditor can trace why the software did what it did. The EU Medical Device Regulation (Regulation (EU) 2017/745) and the EU AI Act's high-risk classification of emotion-inference systems carry the same expectation. HIPAA and 42 CFR Part 2 add that emotional-state data tied to an identifiable patient, especially in substance-use treatment, is protected and must be handled with full provenance. An agent whose emotional behavior is an emergent property of an opaque model cannot satisfy any of these.

The requirement, then, is an emotional substrate that is deterministic, bounded, and auditable by construction, and that modulates the agent's conduct without ever granting it new authority.

2. The Foundation: Affective State as a Governed Control Primitive

This application is built on the affective state field disclosed in United States Patent Application 19/647,395 (Chapter 2), the seventh structural field of the semantic agent schema. Rather than treating emotion as a transient filter on plan selection or a scalar weight on a decision, the disclosed architecture introduces affective state as a deterministic, policy-bounded data structure persisted with the agent across execution cycles, delegation events, and substrate migrations.

Several properties of the disclosed primitive map directly onto the clinical requirements above:

  • Named control fields with defined semantics. The affective state field is organized as a structured modulation layer of named fields, each a measurable axis with defined value ranges, update rules, and governance bounds. The disclosure names, among others, an uncertainty sensitivity field, an ambiguity tolerance field, a novelty appetite field, a persistence-under-partial-failure field, and an escalation-under-time-pressure field. For a therapeutic agent these are exactly the knobs that should respond to a patient's state: elevated uncertainty sensitivity drives more conservative, deferring conduct when a patient's signals are unclear; reduced novelty appetite keeps the agent on validated interaction strategies during a fragile session.

  • Modulation, never authority. A defining guarantee of the disclosed architecture is that affect modulates how the agent weighs alternatives, tolerates ambiguity, and persists under partial failure, but it never creates authority or overrides governance. Every mutation to the affective state field is recorded in lineage, subject to policy validation, and auditable. A therapeutic agent therefore cannot "feel its way" past a safety policy; emotional state shapes tone and deliberation within an envelope that governance defines separately.

  • Policy-bounded updates and decay curves. Each named field carries policy bounds and decay parameters, so no emotional modulation can exceed a policy-defined operating envelope, and transient emotional responses decay rather than accumulating into permanent drift. This is what lets a therapeutic agent register a patient's acute distress in the moment without that distress silently rewiring its baseline behavior afterward.

  • Emotional quarantine. The disclosure describes a volatility detector that computes a composite volatility metric across the recent update history of all named control fields. When volatility exceeds a quarantine threshold the agent enters a restricted execution mode (elevated promotion thresholds, suspended delegation, additional validation), and is released only when volatility falls below a separate, lower recovery threshold, providing hysteresis that prevents oscillatory quarantine cycles. In a therapeutic deployment this is a built-in safety circuit: an agent whose emotional state is being whipsawed by a volatile session degrades into a more cautious, more heavily supervised mode rather than acting confidently on unstable internal signals.

Because affective state lives inside the same governance, lineage, and policy framework as every other agent field, the emotional behavior of a therapeutic agent is reconstructable from the record. That is precisely the property FDA SaMD review, IEC 62304, and ISO 14971 ask for and that emergent-affect systems cannot supply.

3. The Application: Therapeutic and Mental Health Agents

3.1 Tracking emotional trajectory across sessions

Therapy is longitudinal. What matters clinically is not a single utterance but the trajectory of a patient's affect across days and weeks. Because the affective state field is persisted with the agent and updated through a closed feedback loop, structured observations from each session update named control fields under policy bounds, a therapeutic agent maintains a durable, auditable emotional trajectory rather than resetting between conversations. Clinicians can review that trajectory as a structured record instead of inferring it from chat logs.

3.2 Maintaining appropriate therapeutic affect

Effective therapeutic interaction calls for a stable, warm, non-reactive presence. By setting the policy bounds and decay parameters on the agent's own named control fields, a deployment can hold the agent's dispositional baseline inside a clinically appropriate band: bounded responsiveness to provocation, deliberate persistence when a patient disengages, and conservative behavior under uncertainty. The agent's affect responds to the patient without mirroring volatility back at them.

3.3 Safety bounds and crisis handoff

The same governance separation that keeps affect from creating authority is what makes crisis handling safe. Risk-escalation logic, suicidality protocols, and mandatory-reporting paths are governance policy, not emotional state; the affective fields can heighten the agent's sensitivity and bias it toward escalation, but the decision to escalate is enforced by policy the agent cannot modulate away. Emotional quarantine provides a complementary backstop, pulling an agent caught in a destabilizing interaction into a restricted, more heavily validated mode.

4. Deployment Embodiments and Variations

This application supports a range of grounded embodiments, all of which inherit the disclosed governance guarantees:

  • Clinician-supervised adjunct. The agent conducts between-session check-ins; its affective trajectory and any quarantine events surface to a supervising clinician through the auditable lineage record.
  • Substance-use treatment under 42 CFR Part 2. The persisted affective state and its mutation history are handled as protected records with full provenance, satisfying the heightened confidentiality regime while still enabling longitudinal tracking.
  • SaMD-classified mental health software. Named control fields, policy bounds, decay curves, and quarantine lifecycle provide the deterministic, traceable behavior that an IEC 62304 lifecycle and an ISO 14971 risk file require, and that a Predetermined Change Control Plan can describe in concrete, bounded terms.
  • Cross-primitive composition. Per the disclosure, affective state composes with the confidence governor and forecasting engine; a therapeutic agent can let bounded affect modulate its confidence thresholds and its forecasts of session outcomes without that modulation escaping policy control.
  • Delegation within a care team of agents. Where a session is handed between specialized agents, the disclosed affective inheritance and contagion-damping mechanisms bound how one agent's emotional state propagates to another, preventing a single volatile interaction from destabilizing a whole care workflow.

Each embodiment is a faithful implementation of the disclosed affective state primitive applied to a real clinical deployment; none introduces emotional mechanism beyond what the cited application discloses.

5. Why Deterministic Affect Is the Defensible Architecture

The clinical regimes surveyed above do not reject AI that models emotion. They reject emotion that cannot be bounded, traced, or governed. By making affective state a structural, policy-bounded, auditable field that modulates conduct without ever conferring authority, the disclosed architecture turns "the agent has feelings" from a liability into a documented, reviewable engineering property. That is the difference between a therapeutic agent that is conversationally plausible and one that is clinically defensible.

Disclosure Scope

This article describes an application of the affective state primitive disclosed in United States Patent Application 19/647,395. The technology claims, the named control fields, policy-bounded updates, decay curves, emotional quarantine, governance separation, and cross-primitive composition, are grounded in that application. The clinical domain framing, regulatory analysis, and deployment scenarios are application context and are not themselves claims of the cited application. All emotional mechanisms described herein modulate agent behavior and do not create authority or override governance.