Affect-Modulated Promotion Thresholds

Mechanism

The promotion-threshold mechanism inserts a calibrated gating stage between raw affective-state component values and the downstream consumers of those values. Each named component of the affective vector — arousal, valence, motivation, fatigue, social-trust, and any policy-extended fields — is associated with a threshold record. The threshold record specifies a promotion floor below which the component value is suppressed for downstream policy purposes, a promotion ceiling above which the component is treated as saturated, a debounce window that requires the value to remain above the floor for a configured duration before it is admitted, and a hysteresis offset that requires the value to fall a configured amount below the floor before it is demoted.

The mechanism operates as a finite-state element per component. In the unpromoted state, the component's contribution to higher-order policy is bounded at a baseline value irrespective of its instantaneous magnitude. When the instantaneous magnitude crosses the promotion floor, a candidate-promotion timer is initialized. If the magnitude remains above the floor for the entire debounce window, the component transitions to the promoted state, and its instantaneous magnitude is exposed to downstream consumers. If the magnitude falls below the floor before the debounce window completes, the candidate-promotion timer is reset and the component remains unpromoted. In the promoted state, the component is demoted only when the magnitude falls below a hysteresis-offset value strictly less than the promotion floor, ensuring that values oscillating near the floor do not cause repeated promotion and demotion events.

The thresholding stage is purely a gate; it does not alter the underlying affective-state telemetry. The raw component values continue to be recorded in lineage and remain available for audit, diagnostic visualization, and offline calibration. Downstream consumers — the promotion controller for candidate mutations, the delegation authorizer, the planning graph evaluator, the verified-memory commitment pipeline — see only the gated values. The lineage record additionally captures the threshold-state transitions per component, enabling forensic reconstruction of why a particular candidate was or was not admitted at a particular time, and whether the affective state at that time was promoted or merely transient.

Operating Parameters

Each threshold record is parameterized by at least five values. The promotion floor is a scalar in the component's native value space; for normalized components it lies in the open interval between zero and one. The promotion ceiling, when distinct from one, allows policy to treat values above the ceiling as saturated for the purpose of downstream weighting, preventing extreme spikes from disproportionately biasing planning. The debounce window is a duration parameter, typically between one hundred milliseconds and several seconds, that filters transient sensor noise and short-lived perturbations. The hysteresis offset is a non-negative scalar subtracted from the promotion floor to produce the demotion floor; an offset of zero collapses the system to a single threshold and is generally avoided in deployed configurations.

A fifth parameter, the dwell-on-promoted minimum, prevents the system from demoting a freshly-promoted component until a configured minimum dwell time has elapsed even if the magnitude collapses. This is useful for components such as fatigue, where a promoted-fatigue state should produce a measurable behavioral response even if downstream sensor updates briefly contradict the elevated value. The dwell-on-promoted minimum is bounded above by the agent's reactivity envelope to prevent stale affective state from suppressing necessary responses.

Threshold parameters are typed by component and versioned by policy reference. A given agent operates under exactly one policy reference at a time; that reference resolves to a complete threshold table covering every named component in the agent's affective vector. Updates to threshold parameters occur at policy-rollover boundaries, not mid-evaluation, ensuring that any single evaluation cycle observes a consistent threshold surface. Each policy reference, including its threshold table, is recorded in lineage alongside the affective values it gated, so that historical decisions remain reproducible even after parameters are revised.

Calibration of these parameters is performed offline using recorded affective-state telemetry from representative deployment scenarios. Calibration optimizes for two competing criteria: false-promotion rate, the rate at which transient values briefly cross the floor and would erroneously bias policy if admitted, and missed-promotion rate, the rate at which genuinely sustained elevations fail to be admitted within a tolerable latency. The debounce window is the primary control over this trade-off; the hysteresis offset is the primary control over oscillation suppression.

Alternative Embodiments

In a first alternative embodiment, the promotion threshold is a function of the agent's current operating mode rather than a fixed scalar. Mode-conditional thresholds allow an agent operating under elevated caution — for example, after a delegation failure or while traversing a regulatorily restricted region — to require stronger affective evidence before admitting components that would lower selectivity. Mode-conditional thresholds are encoded as a small lookup keyed by the operating-mode identifier, with the lookup result substituted for the static floor and ceiling.

In a second alternative embodiment, the debounce window is itself adaptive, growing during periods of high sensor variance and shrinking during periods of stable telemetry. The adaptive window prevents excessive latency in stable conditions while preserving noise rejection in volatile ones. The adaptation function is bounded above and below by policy constants to prevent runaway adaptation.

In a third alternative embodiment, the hysteresis offset is replaced by a two-sided dead-band centered on the promotion floor. The dead-band embodiment treats values within the band as neither promoting nor demoting, allowing the system to remain in its prior state. This embodiment simplifies parameter calibration in deployments where promotion and demotion symmetry is desirable, at the cost of slightly slower response to genuine state changes.

In a fourth alternative embodiment, the per-component threshold records are replaced by a coupled threshold tensor that accounts for cross-component correlations. For example, elevated arousal coincident with low valence may admit components at a lower magnitude than either component alone would warrant, reflecting the combined affective signal. The tensor embodiment increases calibration complexity but captures interactions that the independent-component embodiment cannot express.

In a fifth alternative embodiment, the thresholding stage exposes a pre-promotion shadow output to a diagnostic channel, allowing observers to see what downstream policy would have done had the gate been disabled. The shadow output is never consumed by the live policy pipeline; it exists solely for calibration, regression testing, and supervised retuning of the threshold parameters.

Composition with Other Primitives

The promotion-threshold stage composes with the named control fields primitive: each control field carries its own threshold record, and the gated output of the threshold stage is what the named-field consumers observe. The threshold stage composes with the candidate-mutation promotion controller by supplying gated affective values that bias the controller's promotion score; the controller treats unpromoted components as contributing only their baseline weighting, which prevents short-lived affective spikes from triggering admissions or rejections that the agent would not endorse over a longer horizon.

The threshold stage composes with the lineage-recording substrate by emitting promotion-state transitions as first-class lineage events. Each transition includes the component identifier, the pre-transition state, the post-transition state, the magnitude that triggered the transition, the policy reference in force, and the wall-clock timestamp. These events allow downstream auditors to reconstruct the affective gating timeline independent of the underlying telemetry.

The threshold stage composes with the confidence-governance primitive without subsuming it. Confidence determines whether the agent is authorized to act at all; the threshold stage determines which affective signals the agent treats as durable enough to influence its selectivity once authorized. The two stages are independent gates and may be tuned independently.

Prior-Art Distinction

Threshold-with-hysteresis gating is a long-established control-systems primitive, found in mechanical thermostats, electrical Schmitt triggers, and signal-processing comparators. The novelty here is not the gating mechanism itself but its placement: between the affective-state estimator and the higher-order policy stages of a cognitive agent, with per-component policy-versioned threshold records and lineage-recorded transitions.

Prior cognitive-agent architectures that incorporate affective state typically expose the affective values directly to policy, either as continuous inputs to a learned function or as scalar weights applied to action utilities. Such architectures are vulnerable to transient noise, sensor outliers, and oscillatory feedback from policy-induced affect changes. The thresholded architecture disclosed here interposes a deterministic, auditable gate that decouples the noise characteristics of the affective estimator from the noise characteristics of the policy stages.

Prior emotion-modeling work in HCI and affective computing has used smoothing, exponential moving averages, and Kalman filtering to denoise affective signals. Those techniques remain available and may be applied upstream of the threshold stage; the threshold stage provides a discrete promotion semantic that smoothing alone does not — namely, a binary promoted/unpromoted distinction that downstream consumers can branch on without reasoning about continuous noise statistics.

Disclosure Scope

This disclosure covers the application of calibrated promotion thresholds, debouncing, and hysteresis to affective-state components prior to their consumption by higher-order policy in a cognitive agent. The disclosure encompasses the per-component threshold record format, the policy-versioning of threshold tables, the finite-state semantics of promoted and unpromoted components, the lineage representation of promotion-state transitions, and the alternative embodiments enumerated above including mode-conditional thresholds, adaptive debounce windows, dead-band hysteresis, coupled threshold tensors, and shadow-output diagnostic channels.

The disclosure is intended to read on any embodiment in which an affective-state component is gated by a calibrated threshold with a debounce window and a hysteresis offset before being admitted to influence agent policy, regardless of the specific functional form of the threshold computation, the encoding of the threshold record, or the choice of downstream consumers. The disclosure is independent of the specific named components in the affective vector and applies equally to canonical fields such as arousal and valence and to policy-extended fields introduced by particular deployments.

The disclosure further extends to deployments in which the threshold mechanism is implemented in a hardware accelerator, a cryptographically attested enclave, a kernel-resident filter, or any other execution environment in which the gate's behavior may be independently audited. It extends to embodiments in which the threshold parameters themselves are derived through automated calibration procedures over recorded telemetry, provided that the deployed parameters are pinned to a policy reference and recorded in lineage at evaluation time. It extends to embodiments in which multiple threshold stages are cascaded, for example with a coarse stage gating admission to a fine stage, and to embodiments in which the threshold record additionally encodes per-consumer overrides that allow distinct downstream consumers to apply distinct gating without changing the underlying telemetry.