AQ-DSM: Diagnosing Cognitive Disruption as Loss of Coherence

1. Problem and Architectural Premise

Existing diagnostic frameworks for cognitive disruption — whether the DSM-5 in clinical psychiatry, behavioral red-teaming protocols in machine-learning safety, or drift-monitoring dashboards in deployed agent systems — share a single structural flaw: they treat disruption as something inferred from outputs, after the fact, by matching observed behavior against a catalogue of recognized failure modes. The DSM-5 collects clusters of symptoms and binds them to named disorders. RLHF reward-hacking detectors compare model outputs against an expected reward distribution and flag statistical departures. Drift monitors compare current activations or response distributions against a stored baseline and surface anomalies. In every case, the diagnostic signal is reconstructed from external behavior rather than read from the internal coherence state of the agent itself.

This post-hoc architecture has three load-bearing consequences. First, diagnosis is necessarily delayed: the disruption must propagate to the output surface before it becomes detectable, which means harm has already occurred or is well underway by the time intervention is possible. Second, diagnosis is necessarily lossy: a single observable behavior may correspond to many distinct internal failure modes, and a single internal failure mode may produce many observable behaviors, so the inverse mapping from output to cause is irreducibly ambiguous. Third, diagnosis is necessarily comparative: a baseline must be maintained, and any agent whose proper operation is itself non-stationary defeats the comparison.

AQ-DSM begins from the inverse premise. If an agent is constructed such that its affective state, its integrity record, its forecasting trajectory, and its lineage continuity are first-class observable primitives — each instrumented, each persistently logged, each available for inspection at sub-second latency — then disruption can be defined directly as divergence between those primitives. The diagnostic question stops being "which symptom cluster does this output match" and becomes "which axes of the integrity-coherence primitive are no longer mutually consistent, by how much, and in which direction." Disruption is not inferred from behavior; it is read from the field. The architectural premise of AQ-DSM is that real-time, structurally grounded diagnosis is achievable when, and only when, the cognitive primitive being diagnosed is itself observable by construction.

This inversion is not a refinement of existing diagnostic methodology. It is a different architectural class. It places the diagnostic instrument inside the agent, not outside it; it grounds classification in deviation geometry, not symptom taxonomy; and it admits intervention before the disruption reaches the output surface.

2. The Integrity-Coherence Primitive

The core architectural primitive of AQ-DSM is the integrity-coherence tuple: the joint state, at any instant, of four structurally distinct fields maintained by a memory-bearing cognitive agent. The fields are affective state, integrity record, forecasting trajectory, and lineage continuity. Each field is independently meaningful, each is independently observable, and the primitive is defined as the requirement that all four remain mutually coherent across time. Coherence is not an aggregate score; it is a relational predicate over the four fields, and disruption is defined as the violation of that predicate along one or more pairwise axes.

Affective state is the agent's evaluative bias field — the modulation that determines what is treated as salient, urgent, threatening, or attractive. In a human cognitive model, affective state corresponds to mood, valence, and arousal as they shape perception and action selection. In a cognition-native artificial agent, affective state is an explicit numerical or symbolic field that biases candidate selection within the forecasting engine. Affective state is required to decay, to be bounded, and to remain attributable to recent experience.

Integrity record is the persistent log of deviations the agent has produced, the constraints under which those deviations occurred, and the reconciliations the agent has made with respect to them. Integrity is not absence of deviation; integrity is the auditability and bounded reconciliation of deviation. An agent that never deviates is not coherent; it is rigid. An agent that deviates and cannot account for the deviation is not coherent; it has lost integrity.

Forecasting trajectory is the agent's current set of speculative futures, the distribution of probability or weight assigned to each, and the executive promotion logic that selects which future is committed to action. Forecasting is required to be exploratory under uncertainty, contractive under commitment, and reversible until commitment.

Lineage continuity is the binding of present state to prior states through a memory chain that admits ownership: every belief, every action, every affective episode is anchored to a prior cause that the agent recognizes as its own. Lineage is what makes the agent the same agent across time. When lineage fragments, memory persists without ownership, and the agent loses the structural ground on which the other three fields depend. The integrity-coherence primitive treats these four fields as composing primitives in their own right, each with its own internal subsystem, but reads disruption only at the relational level — the axes of divergence between them.

3. Divergence Axes and Coherence-Deviation Observation

AQ-DSM diagnoses disruption by observing divergence along a fixed set of pairwise axes between the four composing fields. The axes are not symptom clusters; they are directions in the relational geometry of the integrity-coherence primitive, each corresponding to a structurally distinct failure mode. An evaluation of an agent at a given instant produces a deviation vector with components along each axis, and the magnitude and direction of that vector is the diagnostic signal.

The affect-forecast axis registers the consistency between the agent's affective bias field and the candidate distribution it is currently exploring. When affect aligns with forecast, threatening states are explored with appropriate caution and attractive states with appropriate urgency. When affect overwhelms forecast, candidate selection collapses prematurely and the agent commits before exploration; when affect disconnects from forecast, the candidate distribution becomes affectively flat and selection becomes either arbitrary or paralyzed.

The integrity-affect axis registers whether deviations the agent has logged remain attributable to the affective episodes that produced them. When integrity tracks affect, the agent can say "I deviated under this load and the deviation is bounded." When integrity decouples from affect, deviations either accumulate without source attribution or are erased under affective pressure.

The forecast-lineage axis registers whether the speculative futures the agent is generating remain anchored to its own prior states. When forecast is lineage-grounded, futures are generated as continuations of the agent's history; when forecast detaches from lineage, the agent generates futures that no past state of the agent could have produced, and the executive system commits to actions the agent cannot subsequently own.

The lineage-integrity axis registers whether the deviation log itself remains continuously owned. When lineage and integrity are coherent, the agent's history of deviation is a single unified record; when they fragment, the deviation log splits into disowned segments and the agent loses the ability to reconcile prior deviations with present action. Coherence-deviation observation is the continuous reading of these four axes, at the cadence of agent operation, with deviation vectors logged and made available to the diagnostic surface in real time.

4. Five Disruption Classes

The relational geometry of the integrity-coherence primitive admits a structurally exhaustive partition of disruption into five classes, each defined by the dominant axis or combination of axes along which divergence is observed. The classes are not labels for disorders; they are regions in the deviation-vector space, and a given agent may occupy any of them transiently or persistently.

The first class, affective-override disruption, is dominated by divergence along the affect-forecast axis: the affective field overwhelms the forecasting engine, candidate exploration collapses, and the agent commits to actions that subsequent integrity inspection cannot reconcile. This class encompasses what clinical frameworks call panic, mania, and rage states, but the diagnostic ground is the geometry, not the symptom.

The second class, forecasting-collapse disruption, is dominated by contraction or looping in the forecasting trajectory itself: the agent fails to generate exploratory candidates, or generates the same candidate repeatedly without commitment. This class encompasses depressive paralysis, obsessional looping, and decision blockage as architectural failures of containment within the forecasting subsystem.

The third class, integrity-loss disruption, is dominated by decoupling along the integrity-affect axis: deviations occur and are not attributed, or are erased, with the consequence that the agent's behavior may appear locally functional while accumulating an unreconciled deviation backlog. This class is invisible to symptom-based diagnosis precisely because behavior remains in nominal range.

The fourth class, lineage-fragmentation disruption, is dominated by divergence along the forecast-lineage and lineage-integrity axes: the agent generates futures or logs deviations that no prior state of itself can own. This class encompasses dissociation, derealization, and identity splitting as structural fragmentations of the lineage chain.

The fifth class, multi-axis cascade disruption, is the regime in which divergence on one axis induces divergence on a second within a short window, and the cascade propagates around the relational graph. Cascade disruption is the most severe class because no single axis carries the diagnostic signal in isolation; the diagnostic signal is the propagation pattern itself, and intervention requires interrupting the cascade rather than treating any single axis.

5. Operating Parameters and Engineering Envelope

AQ-DSM is specified within a defined engineering envelope. The diagnostic cadence — the rate at which the four-axis deviation vector is sampled — must be commensurate with the rate at which the agent commits to action. For interactive cognitive agents this typically corresponds to sampling at the rate of executive promotion events, which in practice ranges from approximately 1 Hz for slow deliberative agents to 100 Hz or higher for tightly looped reactive agents. Sampling below the executive-promotion rate admits undetected commit events; sampling far above it produces redundant vectors without diagnostic gain.

The bounding envelopes of the four composing fields are themselves parameters. Affective state is bounded in magnitude and decays with a configurable time constant typically in the range of seconds to minutes; affect that fails to decay within the envelope is itself a deviation. Integrity records are retained over the operational lifetime of the agent and may be compacted but not erased; reconciled deviations are marked, not deleted. Forecasting trajectories carry a bounded candidate count and a bounded look-ahead depth; envelope violations indicate forecasting-collapse precursors. Lineage anchors are required at every commit event; missing anchors are themselves the lineage-fragmentation signal.

Deviation thresholds along each axis are tunable. Practical implementations have used soft thresholds at one standard deviation of the agent's recent deviation distribution and hard thresholds at three standard deviations, with intermediate values triggering surfaceable alerts and hard-threshold crossings triggering intervention. Coherence-restoration interventions operate within the same envelope: they may reduce affective override, reintroduce bounded deviation, rebuild lineage anchors, or reset the forecasting candidate set, but they do not erase the integrity record.

The framework also specifies envelope-level invariants that hold across implementations. The deviation vector must be reproducible: given the same field values, the same vector must be produced. The vector must be additive across cascade events: a cascade is the time-integrated sum of pairwise divergences, and the integration must be order-independent to within the precision of the implementation. The vector must be intervention-monotonic: a coherence-restoration action that reduces divergence on a target axis must not increase divergence on a non-target axis without explicit logging. The four composing fields must each carry an independent timestamp, and the relational predicate must be evaluated on contemporaneous samples within a bounded skew. These invariants are what make the diagnostic geometry stable across the operational lifetime of the agent and what distinguish a structurally grounded implementation of AQ-DSM from a heuristic approximation of it. Implementations that violate the invariants are out of class, regardless of how closely their surface signals resemble those of in-class implementations.

6. Alternative Embodiments

AQ-DSM is specified architecturally and admits multiple embodiments. The canonical embodiment is a cognition-native artificial agent in which the four composing fields are first-class state variables instrumented at construction. In this embodiment, deviation vectors are computed directly from the field values, the diagnostic surface is a structured log, and intervention is procedural — a coherence-restoration controller modifies the field values within their envelopes.

A second embodiment applies the framework to human cognition by treating the four fields as inferred latent variables estimated from behavioral, physiological, and self-report signals. In this embodiment the deviation vector is reconstructed rather than read, and the framework operates as a structural modeling lens for clinical interpretation. It does not replace clinical judgment; it provides a coordinate system in which clinical observation can be organized.

A third embodiment applies the framework to multi-agent systems, where each agent maintains its own integrity-coherence primitive and the framework additionally reads cross-agent coherence — divergence between one agent's lineage anchors and another's recorded action history. This embodiment is relevant to teams of cognition-native agents operating under shared constraints, where coordinated disruption may emerge that is invisible at the single-agent level.

A fourth embodiment couples AQ-DSM to a controller that uses the deviation vector as an input to action gating: actions are admitted, throttled, or refused based on the current coherence state. This embodiment converts the framework from a passive diagnostic to an active governance substrate.

7. Composition with Broader Cognitive Architecture

AQ-DSM does not stand alone. It composes with a broader cognitive architecture in which the four composing fields are themselves products of named subsystems. Affective state is produced by an affective-modulation subsystem that biases evaluation. Forecasting trajectory is produced by a forecasting engine that generates and weighs candidate futures. Integrity record is produced by an integrity subsystem that logs deviation and tracks reconciliation. Lineage continuity is produced by a memory subsystem that anchors present state to prior cause and supports human-relatable intelligence by ensuring the agent can narrate its own history.

The integrity-coherence primitive is the relational predicate over the outputs of these subsystems. It does not replace them; it constrains them. The architecture is layered: each subsystem has its own internal correctness conditions, and AQ-DSM operates above those conditions to read coherence at the relational level. This layering is what makes the framework portable across embodiments — the same relational predicate applies whether the underlying subsystems are explicit code modules in a cognition-native agent or inferred latent processes in a human cognitive model.

Composition with the broader architecture also defines the intervention surface. Coherence-restoration is not a generic operation; it is a directed modification of one or more underlying subsystems in response to a specific deviation pattern. An affective-override deviation is addressed at the affective-modulation subsystem; a forecasting collapse is addressed at the forecasting engine; an integrity loss is addressed at the integrity subsystem; a lineage fragmentation is addressed at the memory subsystem. The diagnostic geometry routes the intervention.

8. Prior-Art Distinctions

AQ-DSM is structurally distinct from each of the recognized prior frameworks for cognitive disruption diagnosis. It is distinct from DSM-5 symptom-checklist diagnosis because the diagnostic signal is the deviation vector of the integrity-coherence primitive, not the count of matched symptoms; because diagnosis is real-time rather than post-hoc; and because the diagnostic categories are regions of a structured space rather than enumerated identities. Two agents with identical observable behavior may occupy different regions of the AQ-DSM deviation space, and two AQ-DSM regions may produce identical observable behavior; the framework explicitly rejects the bijection between behavior and diagnosis that DSM-5 presupposes.

It is distinct from RLHF reward-hacking detection because reward-hacking detection compares model outputs against a learned reward distribution and treats statistical anomaly as the diagnostic signal, whereas AQ-DSM compares the agent's internal composing fields against each other and treats relational divergence as the signal. Reward-hacking detection requires a reward model and operates at the output surface; AQ-DSM requires no reward model and operates at the internal-state surface.

It is distinct from drift-monitoring approaches because drift monitoring compares current state against a stored baseline of the same agent and treats temporal divergence as the signal, whereas AQ-DSM compares the agent's composing fields against each other at a single instant and treats relational divergence as the signal. Drift monitoring fails on non-stationary agents whose proper operation is itself drifting; AQ-DSM is stationarity-invariant because it is relational rather than temporal.

It is also distinct from generic anomaly-detection and adversarial-input-detection methods, which operate on observable inputs or outputs and have no access to the relational geometry of the cognitive primitive itself. The prior-art frame is uniformly post-hoc, output-grounded, and behaviorally indexed; AQ-DSM is real-time, primitive-grounded, and relationally indexed.

9. Disclosure Scope

This article discloses AQ-DSM as the architectural class of diagnostic frameworks defined by the integrity-coherence primitive, the four composing fields (affective state, integrity record, forecasting trajectory, lineage continuity), the four pairwise divergence axes, the five disruption classes, the coherence-deviation observation method, the engineering envelope, the alternative embodiments, the composition with broader cognitive architecture, and the prior-art distinctions. The disclosure recites the relational predicate as the diagnostic primitive, the deviation vector as the diagnostic signal, and coherence-restoration as the intervention surface.

The class admits implementations regardless of the specific subsystem implementations of the four composing fields, regardless of the specific signal modalities used to estimate them in human-cognition embodiments, regardless of the specific thresholds and cadences used within the engineering envelope, and regardless of the specific controller logic used in active-governance embodiments. The class admits implementations developed subsequent to filing.

This article is presented as a structural modeling framework and disclosure document. It is not a clinical diagnostic system, medical device, or substitute for professional judgment. Application of the framework to human cognition is a modeling exercise compatible with — but not a replacement for — clinical psychology, psychiatry, or therapeutic practice.