Affect-Modulated Training Depth

Mechanism

The mechanism couples three subsystems that, in conventional training architectures, operate independently: a curriculum scheduler that sequences training items, a depth-control subsystem that determines how strongly each item is integrated into the agent's parameter substrate, and an affect-measurement subsystem that maintains an internal affective state vector for the agent under training. In the disclosed embodiment, the affect-measurement subsystem produces, at each training step, a vector of scalar fields including at least frustration, uncertainty, curiosity, and confidence. The depth-control subsystem reads this vector and computes an effective depth coefficient that is applied to the gradient update or, in non-gradient embodiments, to the integration weight of the item into a structured substrate.

Concretely, when a curriculum item C is presented at training step t, the system observes the affective response over a short window following presentation. If frustration exceeds a configured threshold and persists across several steps, the depth coefficient for items in the same content class is attenuated by a configured factor, the presentation rate is reduced, or the item is deferred to a later curriculum stage and a placeholder marker is recorded so that later stages can revisit the deferred material when affective conditions permit. Conversely, when curiosity is elevated and confidence is high, the depth coefficient is held at the configured baseline or increased toward a configured ceiling, allowing the agent to consolidate the material more aggressively while internal conditions favor integration.

A critical structural element is the policy bound that prevents the agent from manipulating its own training through strategic affective responses. The affect-to-depth coupling is mediated by a governance policy that constrains the rate and magnitude of depth adjustments, requires multi-step confirmation of affective signals before acting on them, and logs every depth modulation event into a tamper-evident training-governance ledger. This prevents pathological feedback loops in which an agent learns to suppress challenging content by simulating distress, and it preserves auditability of the resulting trained model.

Operating Parameters

Operating parameters of the disclosed mechanism include the affect-sampling cadence, the persistence window required before an affective signal triggers depth modulation, the per-affect threshold values, the depth attenuation factor applied on negative-affect events, the depth amplification factor applied on positive-affect events, the maximum cumulative depth deviation permitted within a curriculum stage, and the deferral horizon for items whose presentation has been postponed.

In one embodiment, affect is sampled at every training step, and a signal must persist across at least three consecutive steps within a configured persistence window before triggering modulation. The depth attenuation factor in such an embodiment ranges from 0.25 to 0.75 of baseline depth, the amplification factor ranges from 1.0 to 1.5 of baseline, and the maximum cumulative deviation within a stage is capped at a configured fraction of total stage capacity to ensure the curriculum cannot be entirely suppressed by sustained negative affect. The deferral horizon may be expressed in curriculum stages, in wall-clock training duration, or in a count of unrelated items that must be successfully integrated before the deferred item is reintroduced.

The mechanism is parameterized to permit operator-specified curves rather than fixed thresholds. A piecewise-linear or sigmoidal mapping may be configured between an affect scalar and the corresponding depth coefficient, allowing operators to tune sensitivity for specific agent populations or curriculum domains. Hysteresis is introduced into the mapping to avoid thrashing when an affect signal hovers near a threshold.

Alternative Embodiments

In a first alternative embodiment, the affect-modulated depth mechanism operates not on a single agent under training but on a population of agents trained in parallel, with population-level affective statistics driving depth modulation for shared curriculum items. Items that consistently provoke negative affect across the population are flagged for curriculum revision rather than being deferred for a single agent.

In a second alternative embodiment, the affect signals driving depth modulation are derived not from an internal affective state vector but from behavioural proxies such as response latency, retry frequency, or output entropy on validation probes interleaved with training. This embodiment is suitable for agent architectures that do not maintain an explicit affective field but exhibit observable correlates of training stress.

In a third alternative embodiment, depth modulation is bidirectional and asymmetric: negative affect triggers immediate depth attenuation with low latency, while positive affect triggers depth amplification only after a sustained confirmation period, encoding a precautionary asymmetry that prefers under-integration over over-integration when uncertainty is present.

In a fourth alternative embodiment, the mechanism is composed with a meta-learning loop in which the depth-modulation policy itself is updated across training runs based on observed downstream agent quality, allowing the governance policy to be refined empirically rather than fixed at design time.

Composition

Affect-modulated training depth is designed to compose with adjacent training-governance primitives disclosed in the broader cognition specification. It composes with curriculum-staging primitives by supplying the deferral signal that staging logic consumes; it composes with lineage-tracking primitives by emitting depth-modulation events into the lineage record so that the trained model's provenance includes its affective training history; and it composes with conformity-attestation primitives by providing verifiable evidence that the training process honoured the disclosed governance policy.

The mechanism further composes with affect-measurement primitives in the affective-state family: any embodiment that produces a sufficiently rich affective vector during training can serve as the input to the depth-modulation logic. This separation of concerns permits independent evolution of affect-measurement techniques and depth-modulation policies.

Composition with pseudonymous-operation primitives is significant because the affective state consumed by depth modulation is, in deployments where pseudonymity is enforced, structurally prevented from leaving the agent's enforcement boundary. The depth-modulation subsystem accordingly runs within the same boundary and writes its decisions, but not the underlying affect values, into the externally visible portion of the training-governance ledger. Auditors verifying the training process learn that depth was attenuated, deferred, or amplified, and on what cumulative schedule, without learning the affective scalar values that drove each individual decision.

The mechanism also composes with confidence-governance primitives at inference time. An agent trained under an affect-modulated depth schedule carries, through its lineage record, evidence that content classes provoking sustained negative affect were integrated only after the affective conditions stabilised; this provenance can in turn be referenced by inference-time confidence governance to calibrate trust in the agent's outputs on those content classes.

Prior-Art Distinction

Curriculum-learning techniques in the prior art adjust the order or difficulty of training items based on loss, accuracy, or competence proxies. The disclosed mechanism is distinguished by its use of an affective state vector, distinct from loss or accuracy, as the input to a depth-control subsystem that modulates parameter-substrate penetration rather than mere item ordering. Prior curriculum-learning systems do not maintain an internal affective field for the agent under training, do not couple such a field to depth control, and do not impose governance bounds on the resulting feedback loop.

Adaptive learning-rate schedulers in the prior art adjust update magnitude based on gradient statistics or validation metrics. The disclosed mechanism differs in that the modulating signal is an internal affective state of the agent, the modulation target is depth of integration into a structured substrate (not merely scalar learning rate), and the policy is constrained by an explicit governance layer that records depth-modulation events into a tamper-evident ledger.

Affective-computing systems in the prior art typically externalise affect estimates for consumption by recommendation engines, tutoring systems, or human supervisors. The disclosed mechanism inverts this arrangement: the affective state is consumed internally as a control signal over the agent's own learning, and external visibility is restricted to the resulting depth-modulation events recorded in the governance ledger. This inversion is structurally significant because it permits affect-aware training to be deployed in settings where externalising affect would be unacceptable on privacy or manipulation grounds.

Disclosure Scope

The disclosure encompasses any training pipeline that conditions depth of integration of curriculum items on a measured affective state of the agent under training, regardless of the specific representation of the affective state, the specific form of the depth-control subsystem, or the specific governance policy bounding the coupling. The disclosure further encompasses systems in which the affective state is replaced by behavioural proxies that serve the same governance function, and systems in which the depth-control subsystem operates on gradient updates, structured substrate weights, or other parameter-integration mechanisms. The scope includes single-agent and population-level embodiments, gradient-based and non-gradient-based training architectures, and embodiments in which the governance policy is fixed, configurable, or itself learned across training runs.