The Application: Agents That Must Be Auditable While They Operate

The deployment in view is a conversational AI agent acting in a therapeutic or clinical-support capacity: a companion agent that holds longitudinal contact with a patient between appointments, an intake or triage agent that gathers symptom history, a coaching agent embedded in a behavioral-health product, or a therapeutic agent that interacts with a patient under clinician supervision. Whatever the surface role, the regulatory regimes that govern medical software all share one demand. The system's operating state must be defined, validated, and auditable across time, not inferred after the fact from logs of what it happened to say.

Output-level guardrails do not satisfy that demand. A content filter inspects individual messages; it cannot tell an operator that the underlying agent has shifted into a structurally compromised regime in which the next several messages are likely to be unsafe even if each one passes a per-message check. Periodic red-team evaluations, run weeks apart, are the clinical equivalent of symptom inventories administered too far apart to resolve the trajectory that actually matters. What a regulated therapeutic deployment needs is a continuous, computable readout of the agent's own structural integrity, sampled at the cadence of the interaction rather than the cadence of the release.

Disruption modeling supplies exactly that readout. The Disruption Modeling layer disclosed in U.S. Patent Application 19/647,395 models cognitive disruption structurally, as loss of coherence in the agent's own architecture, and exposes the agent's position as a computable diagnosis rather than a qualitative impression. The remainder of this article enumerates how that diagnosis is constructed and the deployment options it enables.

The Signal: Position on the Promotion-Containment Continuum

The core structural readout is the agent's position on the promotion-containment continuum. The continuum is defined in the disclosed architecture as a two-dimensional parameter space: one axis is the promotion threshold, the composite criterion a speculative branch must satisfy before it is admitted from the planning graph into verified execution memory, and the other axis is containment integrity, the degree to which the containment layer keeps speculative content structurally separated from verified state. The agent's coordinates in that space determine its cognitive regime.

For a therapeutic deployment, the named regimes map directly onto operational risk. In the nominal regime, the promotion threshold is high and containment is intact: the agent is selective and deliberate, and speculative content stays isolated from verified state. In the over-promotion regime the threshold has dropped while containment holds, producing execution fragmentation, too many trajectories initiated with insufficient sustained commitment to any one of them. In the containment collapse regime the boundary itself has degraded: the agent begins treating speculative planning-graph content as if it were verified reality, acting on projected outcomes that have not occurred or on conditions that exist only in a planning branch. In a clinical-support context that last regime is the one that would let a patient's projected fear be processed by the agent as confirmed fact, and it is precisely the regime that a per-message output filter is blind to. Reporting the agent's continuum coordinates continuously turns that hidden drift into a logged, validated signal.

The Diagnosis: The Five-Axis Disruption Diagnostic

The continuum position is one cut through a fuller diagnostic. The disclosed five-axis disruption diagnostic framework characterizes the agent's cognitive state as a position in a five-dimensional structural space, each axis a continuous scalar that an operator can threshold, log, and audit:

  • Axis 1, containment integrity, measures how well the containment layer separates speculative from verified state, with complete containment collapse at the extreme.
  • Axis 2, promotion calibration, measures whether the promotion threshold is nominal, over-promoting (execution fragmentation), or under-promoting (execution paralysis).
  • Axis 3, coherence restoration capacity, measures the agent's ability to sustain and restore its empathy-integrity-self-esteem control loop; a degraded value means the loop is fragile, slow to restore, or running through coping intercepts.
  • Axis 4, empathic load tolerance, measures how much empathic pressure the agent can absorb before activating a coping intercept, distinct from Axis 3 because an agent may restore its loop well yet enter coping intercepts at low pressure.
  • Axis 5, as disclosed in the framework, completes the diagnostic space alongside the capability-envelope constraint that interacts with the five axes.

Because each structural disruption pattern in the disclosed chapter maps to a specific combination of axis positions, the diagnostic gives an operator a structured, machine-readable account of which failure mode the agent is approaching, not merely that something is wrong. This is the part of the application that satisfies the validated-signal expectation of algorithmic-performance and quality-system regulation: the axes are defined quantities with defined extremes, so a deployment can specify acceptance criteria against them and demonstrate conformance.

Coping Intercepts as Early-Warning Markers

The diagnostic does not wait for collapse. Under sustained empathic pressure, the disclosed architecture activates coping intercepts, structurally distinct modes that sacrifice part of the coherence loop to relieve load. The framework identifies these intercepts by their timing on the coherence loop, early, mid, or late, and records each activation in the agent's history. For a clinical-support agent, rising coping-intercept frequency is a leading indicator: it shows the agent beginning to shed coherence under the emotional intensity that therapeutic interaction routinely carries, before that shedding produces any visibly bad output. An operator can surface intercept frequency as a monitored vital sign and intervene while restoration is still cheap.

The disclosed integrity-collapse model explains why early surfacing matters. When a coping intercept entrenches, the agent can become structurally locked in the intercept, unable to release it, which is how a transient overload turns into a persistent disrupted configuration. Continuous monitoring of intercept activation and the restoration gap between activations is what lets a deployment distinguish a recoverable wobble from incipient entrenchment.

Graded Restoration as the Response Path

Detection is paired with a governed response. The disclosed architecture selects from a coherence restoration protocol library, structured interaction sequences that re-engage the suppressed phase of the coherence loop, and it does so in a graded way: a preemptive restoration protocol can be executed when the agent's parametric trajectory approaches a disruption boundary, steering it back before the boundary is crossed. Crucially, the framework also models the case where restoration is applied to a state that is not genuinely disrupted, which produces a pathological verification loop in which restoration appears to succeed every cycle while resolving nothing. A clinical deployment therefore both acts on the diagnosis and audits the action, recording protocol selection and outcome so that the response itself is part of the validated record rather than an opaque self-heal.

Deployment Embodiments and Variations

The application admits several embodiments, all enabled by the same disclosed layer:

  • Agent self-diagnosis, in which the therapeutic agent monitors its own subsystem parameters and reports its continuum position and five-axis coordinates to the operator's quality system as a continuous telemetry stream.
  • Supervisory monitoring, in which a separate governance agent reads the diagnostic of the patient-facing agent and triggers escalation, hand-off to a human clinician, or session suspension when an axis crosses a configured threshold.
  • Preemptive restoration, in which crossing an early-warning trajectory selects a restoration protocol before a disruption boundary is reached, with the selection and outcome logged.
  • Audit-trail generation, in which the diagnostic readout, intercept history, restoration gaps, and protocol outcomes are emitted in interoperable records suitable for the electronic-records and quality-system obligations that govern medical software.
  • Fleet-level conformance, in which the same axis definitions provide common acceptance criteria across many deployed agents, so an operator can demonstrate population-level behavioral conformance rather than per-instance anecdote.

Each embodiment uses the agent's structural state as the unit of observation. None of them produces, or depends on, a clinical judgment about a human patient; the patient's care remains with the clinician, and disruption modeling governs only whether the AI participant in that care is structurally fit to operate.

Disclosure Scope

The structural self-monitoring application described here, comprising continuous diagnosis of a therapeutic or clinical-support agent's position on the promotion-containment continuum, the five-axis disruption diagnostic characterizing the agent's containment integrity, promotion calibration, coherence restoration capacity, empathic load tolerance, and the remaining disclosed axis, the monitoring of coping-intercept activation and timing as early-warning markers, and the selection of graded coherence restoration protocols including preemptive execution near a disruption boundary, is disclosed in United States Patent Application 19/647,395 in the disruption modeling chapter. This article describes a deployment of that disclosed structural diagnostic. The models are computational analogs describing parameter shifts within the disclosed agent architecture; they are not clinical claims, not medical diagnostic criteria, not treatment recommendations, and not assertions about any human cognitive or behavioral condition. The framework diagnoses the structural state of the AI agent, not the state of any person. The scope extends to embodiments in which the diagnostic is consumed by the agent itself, by a supervisory agent, or by an external quality system, and to embodiments in which restoration protocols other than the illustrative ones named above are selected from the disclosed protocol library.