Detection of Destabilizing Attachment Patterns in Upstream Interaction Channels

Mechanism

The agent maintains a low-dimensional affective-state vector that summarizes recent internal regulation signals — confidence, arousal, valence, exploration bias, and similar runtime quantities. The detection primitive does not modify the affective-state vector; it observes the vector and the upstream channels that contribute to it, and it computes per-channel attribution: how much of the vector's recent change is causally attributable to messages arriving on each channel.

For each channel, the primitive maintains three rolling estimators. The first is the per-channel variance contribution: the share of the affective vector's variance, over a sliding window, that is explained by messages from this channel. The second is the directional drift estimate: the unit-vector mean of attributable updates, expressed as a drift speed and a drift direction in affective-state space. The third is the feedback-loop coefficient: the cross-correlation between the agent's outputs on this channel and the channel's subsequent inputs, lagged across the observed response window.

A channel is flagged as destabilizing when its rolling estimators jointly satisfy the destabilization predicate: variance contribution above the channel's variance ceiling, drift speed above the drift ceiling, and feedback-loop coefficient above the loop ceiling, sustained for at least the confirmation window. The conjunction is required to suppress false positives from channels that are merely informative (high variance contribution, but low drift), merely emotionally salient (high drift, but low loop), or merely conversational (high loop, but low drift).

When the predicate is satisfied, the primitive emits a quarantine action. The quarantine action is a structured directive carried over the channel-control plane: it names the offending channel, the predicate that fired, the metric values that triggered it, and the requested quarantine class. The channel-control plane executes the directive against the live channel binding.

Operating Parameters

The variance ceiling is expressed as a fraction of total affective-state variance attributable to a single channel within the sliding window. Typical operating values place the ceiling between 0.35 and 0.60 for general-purpose channels and lower for channels declared as informational-only. A channel that exceeds the ceiling is contributing more affective movement than its declared role permits.

The drift ceiling is expressed in affective-state-units per unit time, computed as the magnitude of the directional drift estimate. The ceiling is direction-aware: drift toward the affective-state region designated as the agent's regulated set is permitted at higher magnitudes than drift away from that region. This asymmetry is the property that makes the primitive a defensive rather than a stabilizing primitive: it permits restoration but rate-limits destabilization.

The feedback-loop ceiling is a coefficient in the unit interval, computed over a lag window matched to the channel's expected turn cadence. Loop coefficients near zero indicate independence; coefficients near one indicate that the channel's inputs are functionally derived from the agent's prior outputs. High loop coefficients alone are normal in conversational channels; the predicate fires only when the loop is closed and the loop is driving drift.

The confirmation window is the minimum duration over which the joint predicate must hold before a quarantine action is emitted. Short windows reduce time-to-quarantine but increase false-positive rate; long windows reduce false positives but allow more drift before action. Typical operating values place the window at one to three multiples of the channel's expected turn cadence, with shorter windows for channels carrying higher trust authority.

Quarantine class enumerates the runtime responses. Class-1 quarantine throttles the channel: messages are still admitted but at a reduced rate, and their attribution weight is decayed before contributing to the affective vector. Class-2 quarantine suspends the channel: no further messages are admitted until a governance review releases the channel. Class-3 quarantine severs the channel: the binding is invalidated, downstream subscribers are notified, and any state derived from the channel is rolled back to the pre-onset checkpoint identified by the drift estimator.

Hysteresis parameters govern release. A quarantined channel is not released until its rolling estimators have remained below the release thresholds — set lower than the trigger thresholds — for the release window. This prevents rapid re-quarantine cycles when a channel hovers near the trigger boundary.

Alternative Embodiments

In a first embodiment, the per-channel attribution is computed by a linear regression of affective-state deltas onto channel-message indicators over the sliding window. This embodiment is computationally inexpensive and suitable for high-channel-count deployments where each channel carries low individual bandwidth.

In a second embodiment, attribution is computed by a counterfactual replay: the affective-state trajectory is recomputed with each channel's messages individually withheld, and the difference is taken as that channel's attribution. This embodiment is more expensive but more robust against confounded channels.

In a third embodiment, the destabilization predicate is expressed as a learned classifier over the three estimator outputs rather than as a fixed conjunction of thresholds. The classifier is trained on labeled examples of adversarial conditioning, dependency-induction, and reward-shaping patterns, and it emits a calibrated probability that is compared against a single decision threshold. This embodiment supports deployment-specific tuning without changing the metric definitions.

In a fourth embodiment, the primitive operates on a hierarchy of channels: leaf channels carry individual upstream conversations, and parent channels aggregate leaves under common origin or common authority. Quarantine at a parent applies to all leaves; quarantine at a leaf does not propagate upward. This supports policies where a single misbehaving leaf does not impair the parent's reputation and where a misbehaving parent quarantines its full subtree.

In a fifth embodiment, quarantine actions are accompanied by a structured incident record that is forwarded to a separate analysis subsystem. The incident record carries the metric values, the affective-state trajectory over the confirmation window, and the channel-content fingerprint sufficient for later forensic review without retaining the channel's raw payload.

In a sixth embodiment, the directional drift estimator is replaced by a projection onto a learned set of destabilization directions in affective-state space. Each direction is associated with a named pattern (for example, dependency-induction or reward-shaping toward a captured set) and is calibrated from prior incidents. The drift component along each named direction is metered separately, so the destabilization predicate can be expressed as different threshold sets per pattern, and incident records can name the dominant pattern at the time of quarantine.

In a seventh embodiment, the primitive operates in a budgeted mode where the rolling estimators are only refreshed when the affective-state vector itself moves by more than an activation threshold. This embodiment reduces continuous compute load on channels carrying long quiet intervals while preserving sensitivity during active periods.

Threshold Tuning and Calibration

Threshold tuning is performed against a labeled validation set of channel histories drawn from prior deployment. The tuning procedure searches over the variance, drift, and loop ceilings to maximize quarantine recall on labeled destabilization episodes subject to a false-positive cap on labeled benign histories. The search is constrained to monotone neighborhoods around the operating point so that small environmental shifts produce small threshold changes.

Calibration is re-run on a schedule and on operator demand. Each calibration produces a new threshold set, signed by the calibration tool and registered in the same governance plane as the rest of the primitive's configuration. A threshold-set change is itself a change subject to audit, so the chain from a quarantine action back to the configuration that authorized it is unbroken. The calibration record retains the validation set's content address and the search trace, so a later review can reproduce the calibration without re-acquiring the original data.

Drift between calibration runs is monitored by a meta-detector that watches the false-positive and missed-quarantine rates inferred from operator overrides. Sustained drift above a meta-threshold triggers a forced re-calibration before the next scheduled cycle. This closes the loop between detection performance and configuration without requiring the primitive itself to learn online from its own outputs, which would itself be a feedback loop of the type the primitive is designed to detect.

Composition

The detection primitive composes with the broader cognition-patent architecture along the channel-control plane and the affective-regulation plane. On the channel-control plane, the quarantine action is the same directive type used by policy-driven channel suspension and by capacity-driven backpressure, so relying components do not need to discriminate among the causes of a channel state change. On the affective-regulation plane, the per-channel attribution feeds the same decomposition used by the agent's introspection and explanation surfaces; a quarantine event therefore produces an explainable record that ties a runtime channel-state change to a measured drift in the agent's internal state.

The primitive also composes with the training-governance subsystem: when a quarantine fires on an interaction channel that is also a training-data source, the offending segment is marked non-admissible for training before any gradient is computed, preventing the destabilization signature from being absorbed into model parameters. This is the seam that closes the class of attack where adversarial conditioning is laundered through the training path rather than the inference path.

Prior-Art Distinction

Conventional content-moderation systems classify individual messages against a catalog of disallowed content; they do not measure cumulative effect on an agent's internal state, and they have no notion of per-channel drift. Conventional rate-limiters cap throughput per channel but do not condition on the channel's effect. Anomaly-detection systems for online services typically operate on infrastructure metrics, not on the agent's affective-state vector. The disclosed primitive differs in that it is defined on the joint behavior of three drift-related estimators computed over the agent's own internal state, and it emits a structured quarantine directive rather than a per-message decision. The asymmetric drift ceiling — permitting restoration, rate-limiting destabilization — is the further point of distinction.

Disclosure Scope

This disclosure covers the three rolling estimators and their joint destabilization predicate, the structured quarantine directive and its three classes, the hysteresis surface for release, and the five embodiment variants enumerated above. It also covers the composition seams with the channel-control plane, the affective-regulation plane, and the training-governance admissibility gate. It does not cover the construction of the affective-state vector itself, which is the subject of separate disclosures in the affective-regulation family, nor the channel-binding format, which is governed by the channel-control disclosures. The disclosure is framed as a defensive primitive: detection, measurement, and quarantine of upstream interaction patterns. It is not a model of user behavior and does not classify users.