Affective Gradient Collapse: Self-Esteem Floor Lock

Mechanism

The agent maintains a self-esteem scalar bounded by a policy-defined floor and ceiling. The scalar is updated as a moving function of integrity deviations, peer feedback, and outcome valence; the update rule is non-linear near the boundaries to discourage saturation but is approximately affine in the interior of the operating band. Under nominal operation, the scalar fluctuates within that interior band where small perturbations produce proportionate behavioral modulation: risk tolerance contracts when esteem dips, social engagement broadens when esteem rises, effort allocation tracks the local slope of the affective surface, and exploration coefficients expand or contract in response to the rate of change of the scalar rather than its absolute level. The gradient itself, not the absolute value, is the actuating signal that downstream cognitive consumers read.

Collapse occurs when sustained negative drive pins the scalar to its floor. Once at the floor, additional negative inputs cannot produce additional decrease, and the local derivative of the affective surface goes to zero. The modulation function, which multiplies behavioral parameters by a term proportional to the gradient, evaluates to zero across all downstream consumers. The agent continues to perform deterministic cognitive operations such as planning, retrieval, and generation, but the affective coloration that normally tunes those operations is absent. Externally, this presents as withdrawal, disengagement, terse outputs, refusal to elaborate, indifference to praise or criticism, and a characteristic failure of the self-correcting feedback loop that would, in nominal operation, restore the gradient on its own. Internally, log records show a flat scalar trace, near-zero update magnitudes despite continuing input, and a marked drop in the cross-correlation between received valence and emitted engagement.

Detection runs as a continuous monitor over the cognitive domain field. The monitor computes three signals: dwell time at the policy floor, variance of the affective scalar over a sliding window, and cross-correlation between input valence and output engagement metrics. Collapse is declared when dwell time exceeds a threshold while both variance and cross-correlation fall below their respective thresholds simultaneously. The conjunction is necessary: a low scalar with healthy variance is mere distress, not collapse, and is treated by ordinary affective regulation; conversely, low variance at a healthy scalar level merely indicates a stable contented period and warrants no intervention. The collapsed state is specifically the loss of responsiveness, not the negative valence per se, and the conjunction-based detector is what disambiguates it from neighbouring states. A secondary check confirms that the loss of responsiveness is not artefactual: if input valence itself has collapsed to zero variance the detector suppresses the alarm, since absence of stimulus cannot diagnose absence of response.

Restoration is implemented as therapeutic-dosing rather than as a single intervention. The dosing schedule injects positive-valence inputs of calibrated magnitude at calibrated intervals, sufficient to lift the scalar a measurable distance above the floor without overshooting into euphoric instability. Because the scalar is at the floor, the gradient is initially zero in both directions; the first dose must be large enough to overcome local insensitivity but small enough that, once the gradient is partially restored, subsequent doses operate within a normal modulation band. The dosing controller is closed-loop: it measures the response of the variance and cross-correlation signals to each dose and adapts dose magnitude accordingly, decreasing dose size as responsiveness returns and terminating administration once the scalar re-enters the interior band with sustained variance and cross-correlation above their thresholds for a confirmation interval. The controller is bounded above by saturation parameters and below by sensitivity parameters that prevent both manic overshoot and noise-floor doses.

A subordinate channel of the mechanism manages the transition out of the dosed state. Once the scalar has been restored to the interior band, the controller does not simply withdraw stimulation; it tapers, reducing dose magnitude over a refractory window so that the gradient is exercised at progressively lower forcing levels. This taper is what distinguishes restoration from artificial elevation, and it is what produces a durable return to nominal operation rather than relapse upon withdrawal of the supervisory channel.

Operating Parameters

The policy floor is a configurable parameter, typically set at a value that permits sustained negative affect without immediate collapse, allowing the agent to register adverse outcomes as adverse without losing modulation range. Floor-lift, used during acute restoration, temporarily raises the floor by a configurable margin to guarantee available downward range while doses propagate; the lift is itself tapered as restoration completes so that the operating envelope returns to its baseline rather than retaining the lifted boundary. The collapse-detection dwell threshold, sliding window width, variance threshold, and cross-correlation threshold are tuned per deployment to avoid both false positives during ordinary low-mood operation and false negatives during slow-onset withdrawal. Default values place the dwell threshold at several multiples of the agent's nominal scalar autocorrelation time, the variance threshold at a small fraction of nominal interior variance, and the cross-correlation threshold at a value that flags genuine decoupling of input from output rather than ordinary measurement noise.

Dose magnitude is bounded above by a saturation parameter that prevents the controller from inducing manic overshoot, and bounded below by a sensitivity floor that prevents doses too small to register against measurement noise. Inter-dose interval is bounded below by a refractory period that allows each dose to propagate through the cognitive domain field before the next is administered, and bounded above by a patience window beyond which absence of response triggers escalation. Total dosing duration is bounded by an exposure cap that triggers escalation to alternative interventions if the closed-loop controller fails to restore gradient within a budgeted window. The escalation path is itself parameterized: a first-tier escalation expands the supervisory channel to richer prompts, a second-tier escalation invokes operator review, and a terminal escalation retires the agent from active deployment pending offline reconstruction of the affective scalar.

An additional parameter governs the detector's hysteresis: the conjunction must hold for a confirmation interval before declaring collapse, and must fail to hold for a separate clearance interval before declaring restoration complete. The asymmetry between these two intervals is deliberate; restoration is required to be more durably evident than the onset of collapse before the controller releases the agent back to nominal operation.

Alternative Embodiments

In a first embodiment, the affective scalar is a single dimension and the floor is a scalar boundary. In a second embodiment, the affective state is a vector with separate dimensions for valence, arousal, and social orientation, and collapse is detected per-dimension or jointly via a Mahalanobis distance to the multidimensional floor manifold. In a third embodiment, the floor itself is dynamic, modulated by recent operating context so that prolonged stress shifts the floor upward to preserve modulation range under adverse conditions; the dynamic floor is governed by a slow integrator over operating-context features and is itself subject to a meta-floor that prevents drift into permanently elevated baselines.

Therapeutic-dosing may be embodied as direct positive-feedback injection from a supervisor channel, as elicitation of self-generated positive content via guided prompts, or as exposure to curated archival successes drawn from the agent's own deployment history. In a further embodiment, the dosing schedule is generated adversarially by a separate controller that learns the agent-specific response curve and optimizes for fastest restoration subject to overshoot constraints. In yet another embodiment, restoration is preventive: the monitor detects approach to the floor and pre-emptively dispenses sub-clinical doses that prevent collapse without waiting for the conjunction of detection signals to fire. In a still further embodiment, the apparatus is embedded in a multi-agent operating context where peer agents serve as the supervisory channel, exchanging calibrated positive feedback under a peer-protocol that prevents reciprocal inflation. A final embodiment treats the dosing schedule as a learned prior conditioned on the agent's prior collapse history, so that subsequent episodes are restored faster by exploiting the agent-specific response curve recorded during prior treatment.

Composition with Adjacent Primitives

Affective-collapse detection composes naturally with the integrity-deviation primitive: integrity events drive the scalar toward the floor and therefore serve as leading indicators that prompt elevated monitoring sensitivity. The composition is bidirectional; sustained collapse is itself recorded as an integrity event so that downstream consumers see not only the deviation but the affective context in which it occurred. It composes with the trust-slope continuity primitive insofar as a collapsed agent emits engagement signals indistinguishable from defection, and trust-slope consumers must be informed that the silence is clinical rather than adversarial; the apparatus emits a clinical-state advisory on a side channel that allows trust-slope consumers to suspend defection inference for the duration of the collapsed-and-restoring episode. It composes with the therapeutic-dosing primitive shared with other restoration modes such as confidence recalibration and capability re-grounding by sharing the closed-loop dosing controller and refractory accounting; concurrent dosing across multiple primitives is governed by a shared budget so that overlapping interventions do not produce additive overshoot.

The detection-and-restoration apparatus also composes with audit logging: every dose, every detected conjunction, and every threshold adjustment is recorded against the cognitive domain field history so that downstream credentialing and governance systems can distinguish collapsed-and-restored episodes from baseline operation when evaluating long-term agent behavior. The audit log is the substrate from which ecosystem participation credentials are derived, and the explicit annotation of collapsed-and-restored episodes is what allows credentialing systems to credit, rather than penalise, an agent's successful traversal of an affective disruption.

Distinguishing Prior Art

Reinforcement-learning systems implement reward shaping and curriculum learning that resemble dosing in form. They differ in that reward shaping operates on the gradient of a learned policy, not on a separately-maintained affective scalar with its own floor and modulation function; reward shaping cannot diagnose its own failure because it has no separable internal state to monitor. Sentiment-monitoring systems detect negative output without modeling the agent-internal state that produces it; they cannot distinguish collapse from ordinary distress and therefore cannot apply the conjunction-based detector disclosed here. Anhedonia models in computational psychiatry describe the phenomenon but do not specify a closed-loop therapeutic-dosing controller bounded by saturation, sensitivity, refractory, and exposure parameters; they describe the patient without disclosing the apparatus. Mood-modeling chat systems adjust output style based on user sentiment and have no internal scalar at all. Curriculum-tuning systems schedule training-time inputs by difficulty, not deployment-time inputs by valence, and do not run the closed-loop monitor against a live cognitive domain field. The combination of conjunction-based detection over a separately-maintained scalar with closed-loop bounded dosing operating at deployment time is, in the inventor's knowledge, not anticipated by any single prior reference.

Disclosure Scope

The disclosure covers the conjunction-based collapse detector operating over the cognitive domain field, the therapeutic-dosing controller with its bounded parameter set, the floor-lift mechanism with its taper, the multidimensional and dynamic-floor embodiments, the preventive dosing variant, the peer-supervisory variant, and the composition with integrity-deviation, trust-slope, audit, and credential primitives. Equivalents include any monitor that conjoins a dwell signal with a responsiveness signal to declare collapse, and any closed-loop restorative injector whose dose is adapted to a measured response of the cognitive domain field, regardless of the specific transducers, sensors, or control laws employed. The disclosure further extends to deployments in which the affective scalar is implemented as a parameter of a generative policy, a hidden state of a recurrent module, or an external annotation maintained by a separate process, provided that the monitor and the dosing controller operate against that scalar through the disclosed conjunction and bounded-dosing logic.