Affect-Modulated Confidence Sensitivity

Mechanism

The confidence computation function in the cognition patent maps a structured set of input observations to a scalar confidence value through a deterministic evaluation pipeline. Each input carries a polarity, distinguishing observations that argue for higher confidence from those that argue against, and a magnitude expressing the strength of the signal. The baseline computation aggregates these signals through a configured combiner whose gain parameters determine how strongly each polarity is weighted. Affect-modulated sensitivity is the structural mechanism by which a current affective-state declaration is read at evaluation time and used to perturb those gain parameters within bounded, declarative limits.

Affective state, in this architecture, is itself a first-class structured field. It is not a learned latent variable inferred from behavior; it is an explicit declaration drawn from a canonical set such as fear, fatigue, curiosity, calm, and urgency, with each entry carrying an intensity scalar and a provenance record indicating how the state was derived. The mechanism reads this field once per evaluation cycle and applies a configured mapping from the affective declaration to a vector of gain modulators. Each modulator is bounded above and below, and the modulation table is part of the agent's policy reference rather than an internal heuristic.

Three canonical modulations are specified in the patent's reference table. Fear elevates suspicion: the gain on adverse-polarity inputs is increased, so that a signal arguing against confidence is weighted more heavily than it would be under a neutral state. Fatigue raises the registration threshold: the magnitude any input must exceed before contributing to the aggregate is raised, producing a confidence value that responds only to clearer signals and disregards low-amplitude evidence. Curiosity loosens the gain on novelty-polarity inputs: signals carrying a novelty marker contribute more strongly, so that an agent in a curious affective state assigns greater weight to unfamiliar evidence than one in a calm state would.

The bounding constraint is structural rather than aspirational. The modulation table specifies, for each affective entry, the maximum and minimum gain that may be produced. The runtime mechanism applies the modulator and then clips the resulting gain to its declared range. The clipped gain is recorded in the lineage alongside the affective declaration that produced it and the unmodulated baseline gain, so that any downstream auditor can determine, for any historical confidence value, exactly which affective state was in force, which gain perturbation it produced, and whether the perturbation hit its bound.

Operating Parameters

The modulation table is the central operating parameter. Each row of the table binds an affective entry to a vector of gain perturbations indexed by input polarity, with explicit upper and lower bounds for each perturbation. The table is declarative and may be reconfigured per deployment without modifying the mechanism, allowing a clinical-decision agent to adopt one modulation profile and a navigation agent another while the structural pathway remains identical.

A second parameter governs the temporal coupling between affect and confidence. In the simplest configuration the affective declaration is read instantaneously at each evaluation cycle. Alternative configurations apply a smoothing window so that a transient affective spike does not produce a discontinuous gain change. The smoothing window length and shape are declared parameters, and the smoothed affective trajectory is itself recorded in the lineage.

A third parameter controls the priority order when multiple affective entries are simultaneously active. Affective state is not exclusive: an agent may declare both fatigue and urgency at non-zero intensities. The mechanism resolves the joint state through a configured aggregator, typically either a weighted sum of the per-entry modulators or a precedence rule that selects the dominant entry. Both resolutions remain bounded by the per-modulator limits in the table, so that no combination of affective entries can produce a gain outside the declared range.

A fourth parameter governs the floor and ceiling on the resulting confidence value itself. Even when the gain modulators are within their bounds, the aggregated confidence is clipped to a configured envelope before being emitted, ensuring that no affective configuration can drive confidence to zero or to unity unless the underlying evidence independently supports those extremes.

Alternative Embodiments

In a first embodiment the affective declaration is produced by an upstream affective-inference component and consumed unchanged by the modulation mechanism. This embodiment is appropriate where the affective inference is itself a regulated subsystem with its own auditable pipeline.

In a second embodiment the affective declaration is supplied externally, for example by an operator interface or a coupled biological-signal pathway, and the mechanism treats the external declaration as authoritative subject to a freshness check. Stale declarations are reverted to a configured default affective state, ensuring that a disconnected operator interface does not lock the agent into an inappropriate gain configuration.

In a third embodiment the modulation table is parameterized over a context tag, so that the same affective entry produces different gain perturbations in different operational contexts. A fear declaration during a routine reasoning task may produce a modest suspicion increment, while the same declaration during a high-stakes physical action produces a larger one, with both perturbations bounded by their respective table rows.

In a fourth embodiment the gain perturbations are augmented with a structural caveat that propagates with any output produced under non-neutral affect. The caveat does not alter the confidence value but accompanies it through the lineage, so that downstream consumers can distinguish outputs produced under nominal gain from those produced under modulated gain.

Composition With Other Mechanisms

Affect-modulated sensitivity composes with the rest of the confidence-governance pipeline through the gain interface alone. The pipeline's other components, including evidence aggregation, threshold evaluation, and authorization computation, consume the modulated gain without needing to know the affective state that produced it. This separation ensures that the addition or revision of affective entries does not require changes to downstream components.

The mechanism composes with biological-confidence pathways through a shared affective-state reference. When biological signal coherence contributes to the affective declaration, the modulation table reads the resulting state in the ordinary way, with no special pathway connecting biological signals directly to gain modulators. This indirection ensures that the bounded-modulation property holds regardless of the affective declaration's provenance.

The mechanism composes with capability-awareness through the lineage record. A confidence value produced under a fatigue-modulated gain carries that fact into any downstream planner that reads the value, allowing the planner to weigh the confidence against the operational stress under which it was computed. The composition is again deterministic and lineage-recorded, with no opaque coupling between affect and capability.

Prior-Art Distinctions

Existing affective-computing work generally treats affect as an output to be inferred or expressed rather than as a structural input to a computation function. Where affect is fed back into reasoning, it is typically through learned policy networks whose internal coupling between affective representations and downstream decisions is not auditable as a declarative table with explicit bounds.

Confidence-calibration work in machine learning generally operates on the output of a model, applying post-hoc transformations such as temperature scaling or isotonic regression. The mechanism described here operates earlier in the pipeline, on the gain of the aggregator that produces the confidence value, and it does so under a runtime declaration of affective state rather than under a fixed calibration learned at training time.

Bounded-modulation schemes in control engineering, such as gain scheduling, typically index gain on physical operating-point variables. The cognition patent's mechanism extends this pattern to a declared affective state drawn from a canonical set, with the bounding property preserved structurally and the entire modulation made auditable through the agent's lineage record.

Disclosure Scope

This article discloses the structural pathway by which a declared affective state modulates the gain of the confidence computation function under bounded constraints, the canonical fear, fatigue, and curiosity modulations together with their respective effects on adverse, threshold, and novelty pathways, the temporal smoothing and joint-state resolution parameters, and the lineage-recording obligations that accompany every modulated evaluation. The disclosure covers all four embodiments enumerated above and any combination thereof, including embodiments in which the modulation table is parameterized over context tags and embodiments in which structural caveats propagate with modulated outputs.

The disclosure is bounded to mechanisms in which affective state is a first-class declarative field, the modulation table is part of the policy reference, and the per-modulator bounds are enforced at runtime through clipping recorded in the lineage. Implementations that couple affect to confidence through opaque learned policies, that omit per-modulator bounds, or that consume affective state without recording the resulting gain perturbation in an auditable lineage fall outside the disclosed scope.

The disclosure further covers deployments in which the canonical affective set is extended beyond the fear, fatigue, and curiosity entries enumerated above to include additional structured affects such as urgency, calm, frustration, and confidence-in-the-prior. In each case the entry is added as a row in the modulation table with explicit per-polarity gain perturbations and explicit bounds, and the runtime mechanism applies the new entry through the same clipping and lineage-recording pathway used for the canonical entries. The bounding property therefore extends to arbitrary affective vocabularies without requiring a redesign of the mechanism, and the auditability property extends in the same way: any historical confidence value can be reconstructed from the affective declaration, the modulation table version, and the input observations, regardless of which entries the table contains.

The disclosure also covers operator-facing surfaces in which the current affective state, the resulting gain modulation, and the bounded confidence value are exposed for inspection and override. An operator inspecting a confidence value produced under a fear-modulated gain may inject a corrective declaration that is itself a first-class entry in the policy reference, with the override propagating through the standard pathway and the override event recorded in the lineage as a structurally distinguishable input from autonomously declared affective state.