Affect-Modulated Training Depth

Mechanism

Affect-modulated training depth is a further modulation of the depth-selective aggregation mechanism that governs how training content is integrated across the layers of a model. The base mechanism associates each training example with a training depth profile: a per-layer or per-block contribution weight vector that controls the magnitude of the gradient signal from that example permitted to reach each layer or block during the backward pass. Affect modulation adds the affective metadata of training content to the inputs that determine the depth profile. Training examples tagged with high emotional valence, including content associated with safety-critical domains, emotionally sensitive subject matter, or domains in which inappropriate model behavior would cause psychological or physical harm, receive depth profiles specifically tailored to prevent the model from developing uncontrolled deep associations with emotionally charged content.

The affect here is a property of the training content, not an internal emotional state of the model being read back and acted upon. The mechanism does not maintain a frustration, curiosity, or confidence reading for the model under training, and it does not defer items, throttle a presentation rate, or apply thresholds and hysteresis based on the model's reactions. It assigns the shape of a training example's contribution across the depth dimension based on the affective classification of that example's content.

Affective Metadata on Training Content

The affective metadata for each training example is derived from the platform's affect classification infrastructure, which evaluates the emotional valence, emotional intensity, and domain sensitivity of training content as part of the semantic enrichment process applied before training. This affective metadata is incorporated into the training example's semantic metadata alongside the other annotations each example carries: an entropy band classification, a slope position within the trust-slope hierarchy, a content provenance record, and a policy scope. The combined metadata is what the semantic execution substrate evaluates when it determines a depth profile.

Because affective metadata is part of the same semantic enrichment that produces the entropy band and policy scope, affect modulation is not a separate subsystem bolted onto training. It is one more input to the same depth-profile determination that already governs how content from each entropy band is weighted across the model's layers. Content that consists solely of raw text, images, or audio without accompanying semantic metadata cannot be evaluated by the substrate and is inadmissible by default.

Controlled Integration at Intermediate Depths

Training examples with high emotional valence and a safety-critical domain classification receive depth profiles that implement controlled integration at intermediate depths. The controlled-integration profile specifies elevated contribution weights for intermediate layer blocks, the blocks that encode domain-specific knowledge and behavioral patterns, and attenuated contribution weights at the two ends of the depth range. At the shallowest blocks the weights are attenuated because that is where the content's emotional patterns might be triggered by superficial lexical similarity. At the deepest blocks the weights are attenuated because that is where the content's emotional associations might become entangled with the model's most abstract reasoning capabilities.

The intermediate-depth profile is the mechanism by which the model develops structured, domain-contextualized knowledge about emotionally sensitive topics without developing either superficial emotional reactivity or deeply embedded emotional biases. The disclosure does not state numeric weights, thresholds, or layer counts for this profile; it specifies the structural shape, namely elevation at intermediate blocks and attenuation at the shallowest and deepest blocks.

Suppressed Integration of Harmful Content

Training examples associated with content that the platform's governance infrastructure classifies as emotionally manipulative, traumatizing, or psychologically harmful receive suppressed depth profiles that confine the content's influence to the shallowest layers. At those shallow layers the content contributes to the model's lexical and syntactic awareness of the relevant domain without encoding deep conceptual or behavioral patterns.

The purpose of the suppression is twofold. It prevents the model from developing deep associations with harmful emotional content, and at the same time it preserves the model's ability to recognize and respond to references to such content at a surface level. That surface-level recognition is necessary for safety-critical applications in which the model must detect and appropriately handle emotionally charged inputs. Suppression therefore removes the deep encoding while retaining the shallow awareness, rather than excluding the content entirely.

Depth, Not Curriculum Timing

Affect modulation operates on the spatial dimension of training, where in the model a training example is integrated, and not on the temporal dimension of when an example is presented. The temporal sequencing of training content is governed separately by the curriculum engine, which orders examples across training phases. The depth profiles, including the affect-modulated profiles described here, govern the depth at which each example's contribution is encoded. Affect-modulated training depth changes the contribution-weight shape across layer blocks; it does not postpone an example, hold back a content class, or reintroduce deferred material on an affective schedule.

This distinction matters for what the mechanism claims to do. There is no policy in this disclosure that watches the model for distress and slows training in response, and there is no feedback loop in which the model could alter its own training by signaling affect. The affective signal is fixed in the content's metadata before training begins, and it selects a depth-profile shape.

Integration with the Affective-State Architecture

The affect-modulated training depth mechanism integrates with the cross-primitive affective state architecture disclosed elsewhere in the cognition specification. When the platform trains companion agents, therapeutic agents, or other agents that operate in emotionally sensitive contexts, the affect-modulated depth profiles ensure that the agent's training produces structured emotional knowledge, knowledge organized by the same semantic complexity hierarchy that governs all other knowledge in the model, rather than undifferentiated emotional encoding that might produce unpredictable emotional responses at inference time.

This places affect-modulated depth within the same governed-training family as the entropy-band depth profiles, the policy-governed retention and suppression mechanism, and the curriculum-integrated depth scheduling: each contributes a different input to the depth-profile determination, and each is recorded in the training provenance log so that the depth at which any content class was integrated can be audited after training.

Provenance of Affect-Modulated Decisions

Because affect-modulated depth profiles are produced by the same semantic execution substrate that produces all depth profiles, the resulting decisions are recorded in the same training provenance log. For each training example the log records the depth aggregation profile that was applied, the governance record identifying the policy objects that authorized admission and determined the depth profile, and the content provenance record. An auditor reviewing the log can therefore determine that emotionally sensitive content was integrated at intermediate depths, or that harmful content was confined to shallow layers, and on what governance basis, without the depth decision being opaque after the fact.

Disclosure Scope

Affect-modulated training depth, comprising the derivation of affective metadata (emotional valence, emotional intensity, and domain sensitivity) for each training example during semantic enrichment, the use of that metadata as an input to the depth-profile determination performed by the semantic execution substrate, the controlled-integration profile that elevates contribution weights at intermediate layer blocks and attenuates them at the shallowest and deepest blocks for high-valence safety-critical content, the suppressed profile that confines emotionally manipulative, traumatizing, or psychologically harmful content to the shallowest layers while preserving surface-level recognition, and the integration of these profiles with the cross-primitive affective-state architecture and the training provenance log, is disclosed in the cognition filing (U.S. Application No. 19/647,395 and its international counterpart). This article describes that disclosed mechanism. The scope extends to embodiments in which affective metadata is one input among the entropy band, slope position, content provenance, and policy scope that jointly determine a training example's depth profile, and to agent classes, including companion, therapeutic, and embodied agents, trained under affect-modulated depth profiles so that emotional knowledge is organized by the same semantic-complexity hierarchy that governs all other knowledge in the model.