Five-Axis Disruption Diagnostic Framework

Mechanism

The diagnostic framework instruments five independent observation channels, each computing a continuous-valued state estimate from cognitive primitives already produced by the agent's runtime. The promotion-containment axis measures the instantaneous and time-averaged balance between speculative branch promotion events and containment enforcement events. When promotion outpaces containment, the axis registers a positive imbalance proportional to the unguarded speculative volume; when containment dominates, the axis registers a negative imbalance proportional to suppressed inferential capacity. The integrity axis tracks the agent's declared-norm deviation function, which compares the operative behavior policy at time t against the policy declared at session inception, surfacing drift that would otherwise be masked by gradual normative slippage.

The affective axis monitors volatility metrics over the agent's emotional-state primitives, including transition entropy between affective states, persistence half-lives, and oscillation frequency. The confidence axis maintains a calibration ledger comparing the agent's stated confidence against verified outcomes across a sliding window of resolved predictions; calibration error is decomposed into bias and resolution components. The capability axis computes the ratio of utilized capabilities to declared-available capabilities, weighted by the operational relevance of each capability to the active task class — distinguishing benign under-utilization (irrelevant tools left idle) from pathological under-utilization (relevant tools the agent has implicitly disabled).

Each axis emits a numeric severity score on a normalized scale and a categorical state label drawn from an axis-specific taxonomy. The five outputs are assembled into a vector signature that is recorded in the agent's lineage record at every diagnostic checkpoint. Cross-axis interaction matrices, learned offline from labeled disruption episodes, allow the framework to recognize composite syndromes — joint patterns across axes that no single axis would identify in isolation.

The taxonomy underlying the framework distinguishes acute and chronic disruption, primary and secondary axis involvement, and idiopathic versus stimulus-coupled patterns. Acute presentations occur within a short window measured in cognitive cycles and typically reflect a localized perturbation; chronic presentations develop over many checkpoints and reflect drift in the agent's operative configuration. Primary axis involvement identifies the axis on which disruption originated; secondary involvement identifies axes that subsequently drifted as a consequence of the primary disturbance. Stimulus-coupled patterns track temporal correlation between input characteristics and axis trajectories, isolating disruption that the agent's environment is inducing from disruption arising endogenously.

Severity scoring combines the per-axis magnitude with a coupling term that amplifies the score when multiple axes drift in patterns the interaction matrix recognizes as pathological. The composite severity is reported on a band scale (nominal, attentional, advisory, urgent, critical) that drives the cadence and assertiveness of downstream remediation. Course tracking records the sequence of severity bands across checkpoints, supporting clinical-style staging of the disruption episode (onset, progression, plateau, resolution, relapse) and allowing operators to recognize patterns that recur across episodes for the same agent or fleet.

Operating Parameters

The diagnostic cadence is configurable from continuous (every cognitive cycle) to periodic (fixed interval) to triggered (only on suspected disruption signals from the underlying telemetry). Each axis exposes independent thresholds for warning, alert, and critical severity bands, and the bands themselves may be tuned per deployment domain — a high-stakes autonomous controller will use tighter thresholds than an exploratory research agent. The lineage retention window determines how far back trajectory analysis can reach; longer windows enable course-of-disruption tracking but increase storage cost.

Threshold selection follows a calibration procedure in which a reference corpus of labeled episodes is scored under candidate threshold sets and the operating point is selected to balance false-positive and false-negative rates against deployment-specific cost weights. Periodic recalibration is recommended as the agent's operative configuration evolves and as the labeled corpus accrues new episode types. The framework retains earlier thresholds in version-stamped form so that historical findings can be reinterpreted under updated policy, supporting longitudinal studies of agent behavior that span multiple calibration generations without losing comparability.

The interaction matrix that governs composite syndrome recognition is parameterized by a co-activation threshold below which axis correlations are treated as independent noise rather than coupled signal. Confidence calibration uses an adjustable window length and a decay factor that controls how rapidly old calibration evidence is discounted. The capability-axis relevance weighting accepts a task-class taxonomy that maps active tasks to expected tool sets; deployments may supply their own taxonomy or use the default.

Alternative Embodiments

In a first alternative embodiment, the framework operates with a reduced axis set in resource-constrained deployments, omitting the capability and affective axes while retaining promotion-containment, integrity, and confidence — the three axes that most directly bear on safety-relevant disruption. In a second embodiment, additional axes are added: a memory-coherence axis tracking contradiction density across the working memory field, and a goal-stability axis tracking the persistence of declared objectives across reasoning steps. The framework's vector representation generalizes naturally to higher dimensionality.

A further embodiment places the diagnostic computation in a separate process or hardware enclave, isolated from the agent it observes, so that diagnostic findings cannot be corrupted by the agent itself even under adversarial conditions. Yet another embodiment supplies the diagnostic profile to an ensemble of remediation policies, each specialized for a particular composite syndrome, with a routing layer dispatching findings to the appropriate policy. In multi-agent deployments, per-agent profiles are aggregated into a fleet-level dashboard that surfaces correlated disruption — when several agents drift on the same axis simultaneously, an external cause is implicated.

Composition

The framework composes with the upstream coherence-monitoring subsystem, which supplies the raw events the diagnostic axes consume, and with the downstream remediation subsystem, which acts on the findings. Composition with the lineage system is essential: the diagnostic vector at each checkpoint is appended to the lineage record, and trajectory analysis operates on that record to detect monotonic drift, oscillation, or step-change patterns that single-snapshot analysis would miss.

Composition with reinforcement-style learning loops is supported in deployments that adapt the agent's policy from operational experience: the diagnostic vector is supplied to the learning signal as an auxiliary loss term penalizing trajectories that drift toward known pathological syndromes, biasing learned policies toward configurations that the framework scores favorably across all five axes simultaneously rather than optimizing one axis at the others' expense.

The framework composes with audit and explainability tooling: the structured per-axis findings provide a human-legible decomposition of why the agent was flagged, enabling operators to inspect each axis individually rather than confronting an opaque scalar. It composes with policy-update mechanisms that adjust thresholds based on operational experience, and with offline analysis pipelines that mine the lineage corpus for previously unseen syndromes that warrant new entries in the interaction matrix.

Prior-Art Distinction

Prior approaches to agent monitoring typically reduce cognitive state to a single scalar — an alignment score, a confidence measure, or a binary safe/unsafe classification — and conflate distinct disruption patterns under that scalar. An agent exhibiting promotion dominance with intact integrity presents very differently from an agent with promotion dominance and collapsed integrity, yet both appear identical on a one-dimensional promotion metric. The five-axis framework disambiguates these cases by construction: orthogonal axes preserve the structure that scalar reductions destroy.

Prior diagnostic systems also tend to be snapshot-oriented, evaluating the agent at a moment in time without trajectory context. The framework's lineage-coupled trajectory analysis distinguishes acute disruption from chronic drift and identifies course-of-disruption patterns — accelerating, oscillating, plateauing — that inform whether intervention should be immediate, scheduled, or withheld.

Disclosure Scope

Evidence types feeding the framework include direct cognitive-primitive observations (promotion events, containment events, declared-norm comparisons, affective-state transitions, calibration ledger entries, capability invocations), derived statistical summaries (volatility metrics, calibration error decomposition, capability utilization ratios), and structured event traces that preserve the temporal sequence of primitive events for offline analysis. The framework defines admissibility rules for each evidence type — minimum sample sizes for statistical summaries, freshness bounds for primitive observations, integrity requirements for event traces — and reports findings with explicit evidence-quality annotations so that downstream consumers can weight conclusions appropriately.

The disclosure encompasses the five-axis decomposition, the per-axis computation methods, the composite vector representation, the cross-axis interaction matrix, lineage-coupled trajectory analysis, and the integration interfaces with upstream telemetry and downstream remediation. It encompasses embodiments with reduced or expanded axis sets, isolated diagnostic enclaves, and fleet-level aggregation. It encompasses the parameter regimes described above and equivalents thereof, and the application of the framework to any computational agent whose cognitive primitives admit observation along the specified axes, regardless of the underlying model architecture.

The disclosure of U.S. Provisional Application No. 64/049,409 further admits embodiments in which the five-axis vector signature is exported as a portable diagnostic artifact admissible to external review boards, supervisory agents, or cross-fleet correlation services under credentialed exchange, so that disruption findings produced under one operator's diagnostic configuration can be reinterpreted under another operator's threshold policy without requiring re-instrumentation of the originating agent or replay of the underlying telemetry stream.