Emotional Modulation of Planning

Mechanism

The agent maintains an affective-state vector as a first-class canonical field. The vector consists of named scalar components, each component representing a defined affective dimension with a bounded value range. Components are produced by upstream appraisal functions evaluating the agent's recent transition outcomes, lineage signals, and exogenous stimuli; they are not produced by introspection or learned embeddings. Each component is bounded, typed, and decays toward a neutral resting value at a declared rate.

The forecasting engine reads the affective-state vector at the start of each planning round and applies a deterministic modulation function that maps vector components to forecasting parameters. Three principal pathways are defined. The fear component reduces the forecast horizon, contracting the temporal extent over which planning is performed. The curiosity component increases the branching factor, expanding the number of alternative trajectories considered at each planning step. The fatigue component reduces the update rate, lengthening the interval between planning rounds and thereby reducing the rate at which forecasts are republished or recomputed.

The modulation function is total and bounded. For each affective component, the mapping is a monotonic function from the component's value range to a parameter delta, and the parameter delta is clipped to a declared range before being applied. The post-modulation parameter values are themselves bounded, ensuring that no combination of affective inputs can drive the forecasting engine outside its declared operating envelope. The function is stateless: the same affective vector produces the same parameter modulation regardless of history.

Modulation is recorded in lineage. Each planning round captures the affective vector at entry, the parameter deltas applied, and the post-modulation parameter values used. The forecasts produced under modulated parameters are tagged with the modulation record, so that downstream components, including the inference-control layer and any coordinating peers, can interpret the resulting trajectories in light of the affective context that produced them.

Operating Parameters

Each affective component declares a value range, a resting value, and a decay rate. The modulation function declares, for each component-to-parameter pathway, a monotonic mapping and a clipping range for the resulting parameter delta. The post-modulation forecasting parameters declare absolute bounds that are enforced after modulation; these absolute bounds are the same bounds enforced when no modulation is applied, preserving the forecasting engine's structural envelope.

The principal pathways are the fear-to-horizon mapping, the curiosity-to-branching mapping, and the fatigue-to-update-rate mapping. Additional pathways may be declared in the policy reference provided each pathway satisfies the monotonic-and-bounded contract. Pathways are independent; composition across pathways is a parameter-wise sum of the individual deltas, followed by the post-modulation absolute clip.

Alternative Embodiments

In a first embodiment, the affective-state vector is two-dimensional, comprising fear and curiosity components, with a single horizon parameter and a single branching parameter modulated respectively. In a second embodiment, the vector is extended to include fatigue, frustration, and confidence components, each with declared pathways into update rate, decomposition depth, and confidence floor respectively. In a third embodiment, the vector includes a social-affective component that modulates the alignment-scoring weights used in multi-agent coordination, allowing affective state to influence the agent's willingness to align around overlapping forecasts.

Appraisal sources may differ across embodiments. In one configuration, affective components are produced by deterministic appraisal predicates over recent lineage signals, such as the rate of rejection events or the volatility of the trust slope. In another configuration, components are produced by a learned appraisal model whose outputs are clipped and decayed before being exposed as affective values. In a third configuration, components are partially injected by exogenous channels representing operator or environmental signals. In every configuration the downstream contract is identical: bounded, typed, decaying values consumed by the same modulation function.

Modulation may be applied selectively. An embodiment may enable only the fear-to-horizon pathway in a high-stakes operating context, leaving curiosity and fatigue pathways disabled. Selective enablement is a policy-level configuration; the structural mechanism is unchanged.

Composition

The modulation mechanism composes with the forecasting engine's coordination role such that affective state visible to one agent does not propagate uncontrolled into the planning of another. Forecasts published under modulated parameters carry the post-modulation parameters in their metadata but do not carry the underlying affective vector unless explicitly authorized by policy. Receiving agents therefore interpret the structural shape of an incoming forecast, including its horizon and branching, but not the affective state that shaped it.

Composition with the inference-control layer is also bounded. Modulated parameters affect the construction of forecasts but do not alter the admissibility evaluation applied to candidate transitions. An agent in a high-fear state produces shorter, narrower forecasts, but every transition within those forecasts is evaluated against the same admissibility predicates as in a neutral state. Modulation changes what is considered, not what is admitted.

Implementation Considerations

Appraisal predicates are implemented as deterministic functions over recent lineage windows. The fear component, for example, may be implemented as a monotonic function of the recent rejection rate from the inference-control layer combined with a measure of trust slope deterioration. The curiosity component may be implemented as a monotonic function of the novelty of recent committed transitions, computed against an embedding fingerprint of the agent's longer-term semantic history. The fatigue component may be implemented as a monotonic function of cumulative compute consumed within a defined window. Each appraisal is total and bounded; failure modes such as missing inputs produce the resting value rather than undefined behavior.

Decay is implemented as an exponential relaxation toward the resting value, parameterized by a declared half-life. Decay is applied at every planning round entry, before the modulation function reads the vector. The combination of bounded appraisal, bounded decay, and bounded modulation ensures that the affective subsystem is stable in the absence of stimuli: a quiet interval drives every component toward its resting value, which in turn drives the modulation function toward zero delta and restores the forecasting parameters to their unmodulated baselines.

The interaction between modulation and lineage observability deserves note. Because each planning round records the affective vector, the deltas, and the resulting parameters, post-hoc analysis can correlate affective trajectories with downstream outcomes. This enables operators to tune appraisal predicates, decay rates, and pathway mappings based on observed behavior, without modifying any other component of the cognitive architecture. The forecasting engine, the inference-control layer, and the coordination protocol all remain unchanged when affective parameters are retuned.

Safety considerations bound the modulation pathways. Pathways that could increase risk under affective stress, such as a hypothetical pathway from frustration to admissibility threshold, are excluded by construction; the modulation function does not target the inference-control layer's parameters. Affect modulates planning, not admission. This separation ensures that an agent in a degraded affective state produces narrower or slower planning rather than weaker governance.

Prior Art Distinction

Prior approaches to affect in agent systems fall principally into two categories: surface affective expression, in which an agent emits affect-laden output without any internal consequence, and learned affect-conditioned policies, in which affective signals are folded into the policy network as inputs and produce behavior changes whose pathways are not separately inspectable. Neither category provides bounded, declarative pathways from affective state to specific planning parameters, and neither admits independent inspection of the modulation applied in a given round.

The present mechanism is structurally distinct in three respects. First, the affective-state vector is a first-class canonical field with declared components, ranges, and decay rates. Second, the mapping from affective components to forecasting parameters is a bounded, monotonic, declarative function recorded in the policy reference. Third, the modulation is lineage-visible: each round records the vector, the deltas, and the post-modulation parameters, enabling structural audit of how affective state shaped planning. Modulation is bounded by construction; the forecasting engine cannot leave its declared envelope under any affective input.

Boundedness and Stability

Stability of the modulation subsystem is a structural rather than empirical property. Each affective component is bounded by its declared range; each appraisal predicate is total and returns values within that range; decay is a contraction toward the resting value; the modulation function is monotonic and bounded; and the post-modulation forecasting parameters are clipped to declared absolute limits. Under any sequence of appraisal inputs, the forecasting parameters remain within their declared envelope.

Runaway feedback is precluded by the absence of any direct loop from forecasting outcomes back into affective state without passing through bounded appraisal. An adverse forecasting outcome may, through lineage signals, raise the fear component, which narrows the horizon, which produces more conservative subsequent forecasts. The loop is closed but bounded: each step is clipped, and the resting-value relaxation pulls the system back toward neutral in the absence of continued adverse signals. There is no unbounded amplification path.

Tuning stability under operator intervention is preserved because every parameter, range, decay rate, and pathway is declared in the policy reference rather than encoded in component logic. An operator changing a half-life or a clipping bound modifies only the policy document; the affective subsystem rereads the policy at the next round entry and applies the new bounds. There is no model retraining, no warm-up phase, and no risk that an in-flight planning round operates under inconsistent parameters, because parameters are bound atomically at round entry.

Cross-agent stability follows from the policy that affective vectors are not transmitted on the coordination channel by default. Even in deployments where partial transmission is authorized, the receiving agent's appraisal predicates apply the same bounded contract to inbound affective signals as to local ones, preventing one agent's affective state from producing unbounded perturbation in another. The structural envelope of the forecasting engine is therefore robust to both intra-agent and inter-agent affective excursions.

Disclosure Scope

Disclosure encompasses the affective-state vector schema and its component declarations; the appraisal sources and decay semantics that populate the vector; the deterministic modulation function and its declared pathways from components to forecasting parameters; the bounded clipping that preserves the forecasting envelope; and the lineage schema that records each modulation round. The disclosure extends to all component sets satisfying the bounded-and-typed contract, all appraisal sources producing values in the declared ranges, and all pathway configurations satisfying the monotonic-and-bounded contract, including selective and multi-pathway embodiments.