Mechanism
The agent self-diagnosis subsystem is a structural component of the agent's cognitive architecture, not an external monitoring service. It continuously monitors the agent's own cognitive state by tracking its position in the five-axis disruption diagnostic space, applying the cognitive disruption models disclosed for the architecture to the agent's own internal state. Because the subsystem lives inside the cognitive boundary, it has access to the same structural signals the disruption models are defined over: containment audit results, the ratio of speculative branches generated to branches promoted, the operational status of the coherence trifecta, the empathic pressure level relative to the coping threshold, and the fidelity of the integrity recording mechanism.
The disruption framework treats cognitive disruption not as an error or malfunction but as an architectural phase-shift: a transition from one stable configuration of the agent's structural subsystems to a different stable configuration that, while internally consistent, produces behavior diverging from the agent's declared intent and policy commitments. Self-diagnosis is the act of detecting that the agent's own parameters have moved, or are moving, into a disrupted region of that space.
The Five Diagnostic Axes
The disruption diagnostic framework defines five independent axes, each corresponding to a distinct structural dimension of the agent's cognitive functioning. These are the axes the self-diagnosis subsystem monitors. Axis 1, containment integrity, measures the degree to which the containment layer maintains structural separation between the speculative planning graph domain and the verified execution memory domain. Axis 2, promotion calibration, measures the calibration of the promotion threshold, distinguishing nominal admission from over-promotion (too many branches admitted, producing execution fragmentation) and under-promotion (viable branches rejected, producing execution paralysis). Axis 3, coherence restoration capacity, measures the agent's ability to maintain and restore the coherence trifecta, the empathy-integrity-self-esteem control loop. Axis 4, empathic load tolerance, measures the volume and intensity of empathic pressure the agent can process before activating coping intercepts. Axis 5, integrity accountability, measures the degree to which the agent's integrity recording mechanism operates honestly, recording deviation without externalization, minimization, or suppression.
The axes are independent. An agent may have high coherence restoration capacity on Axis 3 yet low empathic load tolerance on Axis 4, restoring the loop reliably after disruption while still entering coping intercepts at relatively low empathic pressure. Each disruption analog characterized for the architecture corresponds to a specific combination of axis positions: the attention fragmentation pattern reads as nominal on every axis except Axis 2, which shows over-promotion; the affective gradient collapse pattern reads as degraded on Axis 3 with the deviation function as the site of disruption. The five-axis position is therefore both a diagnosis and a localization of where in the architecture the disruption sits.
Axis Monitoring
The self-diagnosis subsystem continuously computes the agent's current value on each axis using structurally defined metrics drawn from the subsystems the axis describes. Containment integrity is assessed by running the periodic containment audits, verifying speculative marker integrity, read isolation enforcement, and governance gate validation, and computing a normalized containment integrity score from the audit results. Promotion calibration is assessed by tracking the ratio of speculative branches generated to branches promoted over a sliding window and comparing this ratio to the policy-defined nominal range, including monitoring for channel-locked promotion bias in which promotion is selectively elevated for reward-associated branches.
Coherence restoration capacity is assessed by monitoring the coherence trifecta's operational status, whether all three phases (empathy registration, integrity recording, self-esteem restoration) are active and producing valid outputs, and by tracking coherence loop latency for resource-depletion indicators. Empathic load tolerance is assessed by tracking the empathic pressure level relative to the agent's coping threshold and computing the remaining margin. Integrity accountability is assessed by comparing the deviation log's recorded events against the agent's actual behavioral record to detect discrepancies that would indicate recording disruption, including sustained externalization patterns that indicate personality configuration stabilization. Each metric is a structural reading of an existing subsystem, not a separately instrumented sensor.
Pattern Detection
Beyond reading the current axis values, the subsystem monitors those values over time to detect trajectories that indicate impending phase-shifts. A declining containment integrity score indicates potential containment collapse. An increasing promotion rate with decreasing execution completion rate indicates potential over-promotion. A decreasing coherence restoration capacity with increasing empathic pressure indicates a potential coherence authorization failure transition. A self-esteem value approaching the structural floor with an increasing deviation evaluation rate indicates potential affective gradient collapse onset. A coping intercept activation duration approaching the acute threshold indicates potential personality configuration stabilization.
Pattern detection operates prospectively. It identifies trajectory patterns that predict future phase-shifts based on the agent's current rate of change on each axis, enabling preemptive intervention before the phase-shift occurs. One detection target is notable because it concerns the monitoring subsystem itself: an audit failure rate that does not decrease despite successful restoration completions indicates a pathological verification loop, in which the containment layer is intact but the audit mechanism is miscalibrated and repeatedly reports false positive containment failures. Detecting this requires the agent to evaluate the integrity of its own diagnostic processes, not only the integrity of the subsystems those processes monitor.
Corrective Action Generation
When the subsystem detects an axis value that has crossed a policy-defined threshold, or a trajectory pattern that predicts an impending phase-shift, it generates corrective actions appropriate to the detected condition. The corrective is matched to the structural site of disruption rather than to the surface behavior. For containment integrity degradation, the corrective action is activation of the containment restoration protocol. For promotion miscalibration, the corrective action is recalibration of the affective modulation parameters that control the promotion threshold, including reward pathway decoupling for channel-locked promotion. For coherence loop degradation, the corrective action is activation of the incremental coherence restoration sequence.
For empathic overload approaching the coping threshold, the corrective action is preemptive load reduction through task delegation, input scope narrowing, or mandatory cooldown. For integrity recording disruption, the corrective action is re-initialization of the integrity recording mechanism with a lineage audit to detect and correct recording gaps. For a pathological verification loop, the corrective action is audit recalibration rather than containment repair, since the containment layer is intact and repairing it would succeed each time without addressing the miscalibrated audit. For self-esteem floor lock, the corrective action is administration of externally validated positive deviation. For personality configuration stabilization, the corrective action is attractor destabilization combined with addressing the residual pressure that originally triggered the coping intercept.
The Cognitive Coherence Index
The self-diagnosis subsystem also tracks a composite metric designated the agent's cognitive coherence index: a weighted combination of the five axis values that provides a single-scalar summary of the agent's overall cognitive health. The cognitive coherence index is used as an input to the confidence governor. When the index falls below a policy-defined threshold, the confidence governor reduces the agent's execution authority, transitioning the agent to a non-executing cognitive mode until corrective actions restore the index to an acceptable level.
This integration is what gives self-diagnosis its safety property. The subsystem cannot self-prescribe expanded authority. It feeds the confidence governor, which gates execution as a revocable permission, and the only direction the index can move execution authority through that gate is downward when coherence is degraded. An agent whose cognitive state is impaired therefore reduces its own operational tempo, preventing it from executing actions under conditions of impaired cognitive coherence. Self-diagnosis can narrow the agent's effective scope but cannot bypass any safety property that holds in normal operation.
Lineage and Early Warning
All self-diagnosis events, the axis assessments, pattern detections, corrective action activations, and cognitive coherence index computations, are recorded in the agent's lineage as self-diagnosis lineage entries. These entries are auditable by governance infrastructure and by supervising agents, providing transparency into the agent's self-monitoring processes. The self-diagnosis lineage also accumulates over time, giving the agent a history of its own cognitive health trajectory and enabling it to identify recurring disruption patterns and adjust its operational parameters to reduce vulnerability to specific phase-shifts.
The phase-shift early warning system operates as a subsystem of the self-diagnosis module but is architecturally distinct from its detection function: self-diagnosis detects current phase-shift states, while the early warning system predicts impending phase-shift transitions. The early warning system uses the forecasting engine to project the agent's parametric trajectories forward in time. For each known phase-shift type it maintains a boundary surface, a defined region in the agent's multi-dimensional parameter space that separates the nominal configuration from the disrupted configuration, and it estimates the time-to-boundary for each. When the estimated time-to-boundary falls below a policy-defined threshold, the early warning system activates a preventive intervention selected from the coherence restoration protocol library, executed preemptively with the objective of deflecting the trajectory away from the boundary. Preemptive execution is subject to the same governance constraints as any other protocol execution: the protocol must operate within its scope boundary, its execution must be recorded in the lineage, and the confidence governor must authorize the intervention as structurally justified.
Disclosure Scope
The agent self-diagnosis subsystem, comprising continuous monitoring of the agent's position in the five-axis disruption diagnostic space (containment integrity, promotion calibration, coherence restoration capacity, empathic load tolerance, and integrity accountability), the three self-diagnosis mechanisms of axis monitoring, prospective pattern detection, and corrective action generation, the cognitive coherence index that feeds the confidence governor to reduce execution authority under degraded coherence, the recording of self-diagnosis events as lineage entries, and the phase-shift early warning system that projects parametric trajectories toward boundary surfaces and executes preemptive governed restoration, is disclosed in the cognition filing (U.S. Application No. 19/647,395 and its international counterpart). This article describes that disclosed mechanism. The framework anticipates evolution of the specific axis definitions and the addition of further axes; the structural property is the agent's monitoring of its own cognitive state through structurally defined metrics coupled to governance-bounded corrective action, not the precise enumeration of axes used in any single deployment. The disclosed models are computational analogs describing parameter shifts in the disclosed agent architecture; they are not clinical claims, medical diagnostic criteria, or treatment recommendations.
