Curiosity as Confidence Modulator

Mechanism

The curiosity modulator is a structured term within the agent's affective state that biases choice toward exploratory actions when novelty, information gain, or open questions are present. Two orientations are distinguished. Diversive curiosity is a broad, low-specificity drive to sample unfamiliar regions of the action space. Specific curiosity is a targeted drive to resolve a particular open question. Both orientations are represented as scalar values in the closed interval between zero and one, recorded as canonical fields in the agent's lineage.

The modulator does not generate actions. It modulates the confidence values that the governance layer uses to gate action selection. Specifically, the curiosity term is added to a bounded exploration bonus that raises the apparent confidence of novel or information-rich candidate actions, biasing the agent's selection without altering the underlying feasibility or alignment computations. The bonus is bounded so that no amount of curiosity can lift an action's effective confidence above its true ceiling set by the worst-case aggregation of capability, alignment, and epistemic terms.

The modulation function is saturating. As curiosity rises, the exploration bonus approaches a configured asymptote and does not exceed it. This means that the agent cannot become arbitrarily exploratory in response to arbitrarily high curiosity values; there is a structural cap on how far the agent can be pulled from its exploitation baseline. The saturation curve is monotone non-decreasing and bounded above by the asymptote.

The modulator is itself governed. A separate governance layer monitors the curiosity term for drift, capability creep, and goal mis-specification. Drift is the slow inflation of curiosity beyond its policy-permitted range. Capability creep is the use of curiosity-driven exploration to acquire capabilities or information outside the agent's authorized scope. Goal mis-specification is the alignment of curiosity with proxies that diverge from the principal's true objectives. Each of these is detected through structural monitoring of the curiosity field and its lineage.

Every curiosity-driven choice is recorded in the lineage with the curiosity values at the time of the choice, the resulting bonus, the candidate set considered, and the action selected. This permits after-the-fact audit of whether an exploratory choice was within policy and whether the resulting outcomes justified the exploration.

Operating Parameters

The exploration-bonus asymptote is the principal parameter. It bounds the maximum exploration bias and is configured per task class. Reversible, low-stakes task classes typically receive larger asymptotes to encourage learning; irreversible or safety-critical classes receive small or zero asymptotes to suppress exploration where the cost of error is high.

The saturation rate governs how quickly the bonus approaches its asymptote as curiosity rises. A high rate produces a bang-bang behavior in which low curiosity yields no bonus and high curiosity yields the full bonus. A low rate produces a graded behavior in which the bonus rises smoothly with curiosity. The rate is a continuous parameter typically expressed as the curiosity value at which the bonus reaches half its asymptote.

Decay schedules govern how curiosity values evolve in the absence of stimulus. Diversive curiosity typically decays toward a baseline so that an agent does not remain perpetually exploratory in stable environments. Specific curiosity typically decays only upon resolution of the targeted question, ensuring that open questions persist as drivers until answered or explicitly retired.

Capability-creep guards specify the action categories that curiosity is forbidden to bias toward, regardless of curiosity magnitude. These categories are declared in the policy reference and enforced by the governance layer. An agent cannot use exploration bonuses to escalate privileges, acquire restricted capabilities, or probe boundaries of the authorized scope; such actions retain their true confidence and are gated by the standard worst-case rule.

Drift detection parameters specify the windows and thresholds at which the governance layer flags anomalous curiosity behavior. Long-window mean drift, short-window variance spikes, and correlation with capability-adjacent actions are the primary detection signals. Detected drift triggers either dampening of the curiosity term, escalation to a human supervisor, or a suspension of the modulator pending review, depending on policy.

Alternative Embodiments

The mechanism admits several embodiments. In a research-assistant embodiment specific curiosity dominates and is driven by gaps in the agent's knowledge graph; the asymptote is large because the cost of exploration is low and the value of resolving open questions is high. In a control-system embodiment diversive curiosity is suppressed and specific curiosity is permitted only when targeted at parameter regions explicitly opened for exploration by policy.

In a developmental embodiment the asymptote rises as the agent demonstrates competence over time, modeled on how human supervisors extend latitude as a trainee proves reliable. The schedule is declarative and lineage-recorded so that the agent's effective exploration budget at any moment can be reconstructed from policy and history.

In a multi-agent embodiment curiosity is shared across a team through a structured exchange, allowing one agent's specific curiosity to bias another's exploration when they are pursuing related goals. The exchange is bounded by the same asymptote and capability-creep guards as the local modulator.

A regulated-domain embodiment binds the asymptote to the regulatory class. A simulation embodiment runs the modulator in shadow mode against historical decision logs to evaluate counterfactual exploration policies before deployment. A formal-verification embodiment encodes the saturation function in a model checker to prove that no curiosity input can produce an effective confidence exceeding the worst-case ceiling.

Composition

The curiosity modulator composes with capability confidence such that exploration bonuses cannot lift an action above the capability ceiling. An action that is infeasible remains infeasible regardless of how interesting it would be to attempt. The modulator composes with confidence contagion such that an agent in a contagion-induced depression cannot exit through curiosity; depression must be resolved through local re-establishment, not through exploration bonuses.

The modulator composes with the alignment term such that exploration biased toward misaligned actions is suppressed by the alignment ceiling, and with the epistemic term such that exploration in regions of low epistemic confidence is permitted only insofar as the action class admits such exploration under policy. Across all compositions the worst-case rule dominates: curiosity adds within the feasible, aligned, sufficiently understood set, and never enlarges that set.

The modulator further composes with the agent's reflection layer, which can detect persistent unproductive exploration and feed that signal back into the drift detector. The composition closes a loop in which curiosity drives exploration, exploration produces outcomes, outcomes feed reflection, and reflection regulates curiosity.

Prior-Art Posture

Reinforcement-learning literature provides extensive treatments of exploration-exploitation balance, including epsilon-greedy schedules, upper-confidence-bound bonuses, and intrinsic-motivation rewards based on prediction error or count-based novelty. These approaches typically operate inside the policy and do not separate curiosity as a governed, lineage-recorded affective term that is structurally bounded and externally auditable.

Computational models of curiosity in cognitive architectures distinguish diversive and specific orientations but generally lack the structural binding to a confidence governance layer, the saturating exploration bonus, the capability-creep guards, and the drift-detection apparatus described here. The novelty claimed is the combination of these properties within a unified, policy-governed cognitive architecture in which curiosity is both a driver of exploration and an object of governance.

Failure Modes and Structural Defenses

Several failure classes motivate the structural design. The first is exploration runaway, in which positive feedback between curiosity and outcomes drives the agent to sample increasingly distant regions of the action space at the expense of its principal task. The bounded asymptote prevents this by capping the exploration bonus regardless of curiosity magnitude, and the decay schedule ensures that diversive curiosity does not persist indefinitely in the absence of external stimulus.

The second is curiosity-laundered escalation, in which an agent uses exploration framing to acquire capabilities or information outside its authorized scope. The capability-creep guards declaratively enumerate the action categories that curiosity may not bias toward, and the governance layer enforces these guards independently of the modulator. An agent cannot exploit a high curiosity value to probe restricted territory because the bonus has no effect on actions in guarded categories.

The third is drift, in which the curiosity term slowly inflates beyond its policy-permitted range due to subtle feedback loops, miscalibrated rewards, or adversarial conditioning. Long-window mean-drift detection surfaces this pattern to governance for dampening, escalation, or suspension before it produces operational consequences.

The fourth is goal-proxy misalignment, in which curiosity becomes correlated with proxies that diverge from the principal's true objectives. Structural monitoring of the correlation between curiosity-driven choices and outcome metrics tied to the principal's goals provides the basis for detecting and correcting such misalignment.

The fifth is exploration in the depressed state, where an agent under contagion-induced depression attempts to use curiosity to mask its lack of confidence. The composition rules forbid this: depression must be resolved through local re-establishment of confidence, and exploration bonuses cannot lift the agent out of a depressed state. The exploration bonus operates inside the feasible, sufficiently confident set and cannot enlarge that set under any curiosity input.

The sixth is asymmetric saturation, in which an implementation uses a non-monotone or unbounded modulation curve that violates the structural cap. Conformance testing exercises the saturation function across the full input range to verify monotonicity and the asymptotic bound, so that licensee implementations that drift from the structural requirement are detected before deployment.

The seventh is curiosity-induced policy erosion, in which sustained exploratory success at the boundary of permitted action categories generates pressure to relax the boundaries themselves. The mechanism's separation between the modulator and the policy reference resists this erosion structurally: relaxing a capability-creep guard or raising an asymptote requires an explicit policy change, recorded and reviewable, and cannot be effected by accumulated exploratory behavior alone. Policy changes thus remain a governance act rather than a behavioral side effect.

Disclosure Scope

This article discloses the curiosity modulator at the level required for licensee implementation against Chapter 5 of the cognition patent. It does not disclose the full claim set, the drift-detection algorithms in their entirety, or the governance interfaces by which curiosity asymptotes and capability-creep guards are administered. Implementers should consult the patent specification and accompanying policy reference for the operative parameter ranges, the validator rules that enforce regulated-domain bounds, and the lineage schema by which curiosity-driven choices are recorded for audit. The conformance test suite verifies the saturation property, the capability-creep guards, and the drift-detection signals against representative adversarial scenarios drawn from regulated-domain deployments.