Affect as Cross-Primitive Input

Mechanism

The affective state of a cognitive agent is represented as a structured field over a fixed dimensional schema — for example, valence, arousal, confidence-tone, exploration-tone, caution-tone, and frustration — each dimension carrying a magnitude and a recency-weighted decay. The field is generated by the affective-state primitive from inputs supplied by other cognitive primitives: integrity-coherence reports its coherence residuals; capability-awareness reports successful and failed execution outcomes; forecasting reports the surprise between projected and actual outcomes; semantic discovery reports discovery-intent satisfaction. The affective primitive integrates these inputs over a configurable temporal window, producing the current affective field as a continuously updated state.

The architectural distinction of the disclosed mechanism is that the affective field, once produced, is a published input available to other cognitive primitives by reference. The confidence-computation function, when invoked to produce a confidence estimate over a candidate execution, reads the affective field as one of its parameters; specific dimensions modulate the gain applied to capability inputs (so that a frustrated agent computes lower confidence at the same nominal capability), to knowledge inputs (so that a cautious agent demands a higher knowledge bar before issuing a confident estimate), and to environmental inputs (so that an aroused agent weights environmental signals more heavily). The forecasting engine, when invoked to project the consequences of a candidate, reads the affective field to set its branching factor (a confident agent speculates more broadly), its pruning aggressiveness (a cautious agent prunes earlier), its speculative depth (an exploratory agent looks further ahead), and its delegation preference (a frustrated agent more readily delegates to peer cognitive resources). Semantic discovery, where present in the agent's invocation graph, reads the affective field to set the weighting between immediate-neighbor and forecast-outcome scoring described elsewhere in this disclosure.

The coupling is structurally read-only. The consuming primitives observe the affective field; they do not mutate it. The only path from a consuming primitive's behavior back into the affective field is through the affective-state primitive's input collection, which observes the consuming primitive's outcomes and integrates them into the next field update. This indirection is essential to the stability of the composed system: a direct write-back from a consuming primitive into affect would create an uncontrolled feedback loop; the indirection through outcome-observation imposes a deliberate update cadence and an explicit integration function that can be tuned for stability.

Operating Parameters

The coupling weights — that is, the gains by which each affective dimension influences each consuming primitive's parameters — are specified in the agent's policy reference and are independent for each (dimension, primitive) pair. A policy may, for example, couple frustration strongly to confidence-gain modulation but weakly to forecasting branching-factor modulation, producing an agent whose execution willingness is sensitive to recent failure but whose planning breadth is comparatively stable. The weight space is therefore high-dimensional, and the disclosure contemplates that practical agent designs will publish weight presets corresponding to behavioral archetypes — exploratory, conservative, deliberative, reactive — together with the policy-level controls to depart from them.

The temporal integration window over which the affective primitive accumulates outcome inputs is a further operating parameter and trades responsiveness against stability. Short windows (seconds to minutes) produce affect that tracks recent events closely, suitable for fast-iterating tasks; long windows (hours to operational sessions) produce affect that reflects sustained experience, suitable for complex tasks where transient failures should not disrupt strategy. The architecture admits multiple integration timescales operating concurrently — fast and slow affect — both published into the field and each consumed at separately specified weights by the downstream primitives. The decay constants, integration kernels, and dimensional coupling matrices are all policy-published and runtime-observable.

Alternative Embodiments

The dimensional schema of the affective field is itself a parameter of the disclosure. A minimal schema may use a single valence dimension; a richer schema may include separate dimensions for each archetypal affective tone; an expansive schema may include domain-specific dimensions (a clinical-decision agent may carry a separate caution-tone for patient-safety contexts independent of its general caution-tone). The architecture imposes no upper bound on dimensionality; it requires only that consuming primitives declare which dimensions they read and that the policy reference supply weights for the declared couplings.

Alternative embodiments differ in the set of consuming primitives. The minimal coupled embodiment couples affect to confidence only; the canonical embodiment couples affect to confidence and forecasting; the expansive embodiment couples affect to confidence, forecasting, semantic discovery, execution gating, and inter-agent delegation. The disclosure contemplates all such couplings as instances of the same cross-primitive-input mechanism, with the boundary of the coupling fixed by the agent's declared invocation graph rather than by the affective-state primitive itself.

Embodiments also differ in the input pathway by which affect is generated. A direct embodiment integrates outcome signals from the consuming primitives only; an enriched embodiment additionally accepts exogenous affective inputs from communicating peer agents (so that affect can be socially propagated in a multi-agent system) or from a human operator (so that affect can be commanded for safety-critical override). Each enrichment is governed by the same policy reference and is subject to the same read-only discipline imposed on consuming primitives.

Composition

Cross-primitive affective input is the principal mechanism by which the disclosed architecture produces coherent agents. Without it, each cognitive primitive would respond only to its own immediate inputs, and an agent's behavior would decompose into independent subsystem behaviors that share no experiential context. With it, experience accumulated through any one subsystem propagates through affect to all subsystems, producing the gestalt behavioral profile characteristic of an experienced agent: caution learned through failed execution propagates to more conservative planning and stricter confidence; exploratory success propagates to broader speculation and more permissive discovery; frustration with one approach propagates to an increased delegation preference across all approaches.

The composition is itself bounded. Because affect is consumed read-only, there is no path by which a confidence shortfall directly suppresses future confidence: it is mediated through outcome observation and affective integration, which can be tuned, observed, and bounded. The composition therefore admits formal stability analysis under the policy weights, and the disclosure contemplates that practical deployments will include affect-stability monitoring as an operational concern alongside the standard cognitive-primitive observability.

Worked Example

Consider a clinical-decision-support agent operating across a sequence of patient encounters. In an early encounter, the agent issues a confident recommendation that, in retrospect, proved suboptimal: the integrity-coherence primitive flags the post-hoc inconsistency, the capability-awareness primitive records the failed-execution outcome, and the forecasting primitive registers the surprise between projected and actual outcomes. The affective primitive integrates these inputs into the affective field, increasing the caution-tone dimension and decreasing the confidence-tone dimension over the integration window.

In the next encounter, the consuming primitives observe the updated affective field. Confidence computation, weighting capability inputs through a caution-attenuated gain, returns a more conservative confidence estimate over the same nominal capability profile. Forecasting, with caution-modulated branching, explores a broader set of differential considerations before committing to a recommendation. Semantic discovery, weighting forecast outcomes more heavily under elevated caution, navigates more deeply through corroborating literature anchors before surfacing supporting evidence. The composite agent behavior is not merely "less confident" — it is qualitatively more deliberative across every cognitive subsystem, and the deliberation is directly traceable to the experiential signal generated by the prior failed execution.

Over a longer horizon, sustained successful execution allows the slow-affect integration to recover the confidence-tone dimension, and the agent's behavior across all subsystems progressively returns to its baseline operating posture. The fast and slow integrations admit transient reactions to recent events without permanently distorting the agent's behavior, and the slow integration ensures that durable experiential learning is preserved across sessions. This example illustrates the value of cross-primitive coupling: an agent without it would carry the failed-execution signal in capability-awareness alone, where it would influence subsequent confidence at that specific capability but would not generalize to broader deliberation, planning depth, or evidence-gathering thoroughness.

Prior Art Distinction

Affective computing systems that model emotion have long existed; they typically generate affect as an output for human-facing display, not as a structured input to further cognitive subsystems within the same agent. Reinforcement-learning agents implicitly modulate behavior through value-function updates from experience but do so without an externalized, observable affective field consumed under explicit read-only discipline by separately specified primitives. Cognitive architectures from the symbolic tradition occasionally include emotion-like state but rarely formalize it as a published field with declared cross-primitive couplings.

The disclosed mechanism is distinguished by the explicit publication of affect as a field consumed by multiple separately specified cognitive primitives, by the read-only discipline that prevents direct write-back, by the policy-published weight space governing each coupling independently, and by the multi-timescale integration that admits coherent fast and slow affective behavior in the same agent.

Stability and Governance

Because the disclosed mechanism creates feedback loops between cumulative experience and downstream cognitive behavior, stability of the composed system is a first-class architectural concern. A naively configured agent could enter pathological regimes — cascading caution that suppresses execution, runaway exploration that fails to commit, oscillatory affect that produces erratic behavior — and the architecture provides explicit means to prevent and detect these regimes. The read-only consumption discipline ensures that no consuming primitive can directly amplify its own affective input; the temporal integration window in the affective primitive imposes a deliberate update cadence that bounds the rate at which affect can change; the policy-published coupling weights are bounded above and below to prevent any single dimension from dominating; and the multi-timescale integration ensures that fast affective reactions cannot persistently distort the slower experiential record without genuine sustained input.

Operational governance of the affective subsystem proceeds through the same observability infrastructure that the broader cognition architecture provides. Each affective field update is logged with its constituent inputs, integration kernel state, and resulting field values; each consuming-primitive invocation publishes the affective field values it read and the parameters it derived from them. Operators can therefore audit, in retrospect, the affective trajectory of an agent over a session and attribute downstream behavioral choices to specific affective inputs. Where formal stability guarantees are required — as in safety-critical deployments — the architecture supports operating modes in which affective coupling is bounded to validated weight presets, in which exogenous human-operator affective override is available, and in which affective excursions outside specified envelopes trigger automatic fallback to a baseline operating posture.

Disclosure Scope

The disclosure encompasses the structured affective field as a published input to multiple cognitive primitives; the read-only consumption discipline imposed on consuming primitives; the policy-published weight space governing the coupling between each affective dimension and each consuming-primitive parameter; the multi-timescale integration that admits concurrent fast and slow affective fields; the alternative dimensional schemas, alternative consumer sets, and alternative input pathways enumerated above; and the principle that cross-primitive affective input is the architectural mechanism by which an agent's experience in any one cognitive subsystem produces coherent modulation of all its cognitive subsystems.