Mechanism

The governed training loop reconceives a training iteration as a proposed semantic mutation to the model's accumulated knowledge state. The same semantic execution substrate that governs inference-time behavior and discovery traversal is positioned at the training-loop boundaries, so that the architectural principles that govern what a model may say at inference time also govern what a model may learn at training time, through the same substrate, the same policy objects, and the same provenance infrastructure. The substrate operates at the boundary between the forward-pass loss computation and the backward-pass gradient application: gradients are computed as in conventional training, but the gradient signal is modulated, gated, or selectively routed across model depth based on an admissibility determination produced by the substrate.

The substrate does not alter the mathematical machinery of gradient computation or optimizer updates. It governs which gradient signals reach which layers and with what magnitude, based on the semantic properties of the training content that produced those gradients. Just as inference-time governance intercepts candidate inference transitions before commitment, training-time governance intercepts candidate parameter updates at the point where the training pipeline produces them, evaluates them for admissibility, and records a provenance trail of every decision: every admission, every rejection, every modulation. A training batch provides input to the substrate, which feeds a depth profile router that directs the modulated gradient signal toward shallow layers, middle layers, or deep layers of the model.

Training Examples as Proposed Semantic Mutations

Each training example presented to the model is treated as a proposed semantic mutation to the model's learned representations: a structured modification to a governed state that must be evaluated for admissibility before commitment. This treatment produces three architectural consequences. First, non-training, the refusal to integrate a training example into the model's parameters, is a valid computational result rather than an error condition. The substrate may determine that an example is inadmissible because its policy constraints prohibit integration, because its entropy characteristics are incompatible with the current training phase, or because its provenance metadata fails validation, and may reject the example, producing an iteration in which the parameters are not updated. The non-training result is recorded in the training provenance log as a governed event.

Second, each training example must carry semantic metadata sufficient for the substrate to render a determination. An example consisting solely of raw content without accompanying metadata cannot be evaluated and is inadmissible by default. The required metadata comprises at minimum an entropy band classification indicating the semantic complexity and information density of the content, a slope position indicating the content's position within the platform's trust-slope hierarchy, a content provenance record identifying the source and chain of custody, and a policy scope identifying the governance constraints that apply. Third, the training corpus becomes a governed collection of semantically annotated objects, each carrying its own governance profile, rather than an undifferentiated mass of data. The substrate's evaluation is not limited to binary admit or reject determinations: it may render graded determinations that specify how an example's contribution should be weighted, routed, and distributed across the model's depth.

Entropy-Band-Indexed Training Depth Profiles

Each entropy band recognized by the platform's entropy extraction pipeline is associated with a training depth profile that governs how content from that band is selectively weighted across the layers of the model during training. The entropy band classification is derived from the semantic entropy of each example, the information-theoretic divergence of the example's semantic embedding distribution relative to the model's current representational state. Content with low semantic entropy, content well-represented in the model's existing knowledge, receives shallow depth profiles. Content with high semantic entropy, content introducing novel semantic structure, receives deep depth profiles.

The training depth profile is a structured data object comprising a per-layer or per-block contribution weight vector. For a model comprising L layers or B blocks, the profile specifies a weight value for each layer or block governing the magnitude of the gradient signal from the associated example that is permitted to influence that layer or block. A weight of one permits the full gradient signal to reach the layer; a weight of zero prevents any gradient signal from reaching it; a weight between zero and one attenuates the signal; a weight greater than one amplifies it. The association between entropy bands and depth profiles is not fixed prior to training: a profile adaptation engine monitors the model's internal entropy distribution at defined checkpoints and adjusts the profiles to maintain alignment between the entropy band structure of the corpus and the entropy structure of the model's internal representations. High-entropy content is increasingly directed toward deep layers where multi-step abstraction occurs, and low-entropy content is increasingly confined to shallow layers where local pattern matching occurs, preserving deep representational capacity for content that requires it.

Depth-Selective Aggregation Mechanics

Depth-selective aggregation is the mechanism by which the training depth profiles are applied to the gradient signal to produce content-governed parameter updates. It operates at each layer transition during the backward pass, modulating the gradient signal for each example according to the depth profile associated with that example's entropy band. This is distinguished from layer-wise aggregation developed for federated learning, which assigns different aggregation weights to different layers across multiple model instances being merged. The present mechanism does not aggregate multiple models; it governs the depth at which a single example's gradient contribution is integrated into a single model's parameters, based on the semantic properties of the content that produced the gradient.

The mechanism is implemented through one or more of three complementary techniques. The gated residual connection technique augments each residual connection with a gating coefficient derived from the depth profile, multiplying the gradient signal flowing through the residual pathway during the backward pass while leaving the forward pass unchanged. The attention-based depth selection technique supplies the depth-profile weight as an additional input to the attention computation, scaling the gradient that reaches the attention weights and value projections during backpropagation. The layer-specific scaling factor technique applies a scalar multiplier to the gradient signal at each layer boundary during the backward pass and is architecture-agnostic, requiring only the ability to intercept and scale the gradient at each boundary. To keep profile evaluation tractable in deep networks, the mechanism operates at block-level granularity, grouping contiguous layers into blocks and specifying per-block weights. The mechanism does not alter the optimizer's update rule: it alters the gradient signal the optimizer receives, and is compatible with stochastic gradient descent, Adam, AdamW, and their variants.

Policy-Governed Knowledge Retention and Suppression

The depth-selective aggregation mechanism is integrated with the platform's rights-grade content governance so that the same policy objects that govern agent behavior and inference-time admissibility govern the depth and magnitude of knowledge integration during training. Content admitted under time-limited licensing agreements is trained with a suppressed depth profile, in which the contribution weights for the model's deeper layers are set to zero or near-zero, confining the content's influence to shallower layers. Content that may need to be de-emphasized or removed in the future should not be deeply encoded, where removal would require invasive procedures. Content from the governed exclusion corpus is structurally prevented from integration through a zero-weight depth profile that sets the contribution weight to zero at every layer.

The distinction from post-hoc unlearning is structural. Post-hoc unlearning operates after training, approximating the influence of content on the parameters and applying corrective updates intended to remove that influence. It is inherently approximate because the influence of any single example is diffused across the parameters through the non-linear dynamics of gradient-based optimization. The present disclosure implements structural prevention instead: content whose governance profile restricts deep integration is prevented from deep integration at training time, before the gradient signal reaches the deep layers. There is no need to unlearn what was never deeply learned. The prevention is deterministic, auditable, and reversible by changing the depth profile, whereas post-hoc unlearning is stochastic, approximate, and irreversible. When multiple policies apply to a single example, the substrate resolves the applicable profile by applying the most restrictive policy, and the resolution is recorded in the training provenance log.

Provenance-Traceable Training Dynamics

The substrate records a comprehensive provenance trail for every training iteration, producing a training provenance log that is the training-time analog of the lineage field maintained for semantic agents, inference processes, and discovery traversals. For each batch or example, the log records the entropy band classification, the slope position, the depth aggregation profile that was applied, the per-layer contribution weight that actually reached each block after modulation, a governance record identifying the policy objects that authorized admission and determined the profile, a content provenance record, and an admissibility determination record indicating whether the example was admitted, rejected, or admitted with a modified profile. The log is chronologically ordered and append-only, with each entry timestamped, sequentially numbered, and annotated with the epoch, iteration, and batch index, making it tamper-resistant.

The log supports post-training provenance queries that reconstruct influence pathways between training content and model capabilities. A forward query begins with an example or class of content and traces the depth profile, contribution weights, and governance decisions that governed its integration. A reverse query begins with a model behavior observed at inference time and traces backward to identify the training content whose depth profiles encompassed the layer blocks active during the behavior. The reverse query does not definitively attribute behavior to specific content, because the non-linear dynamics of optimization preclude exact attribution, but it identifies the set of content that was structurally permitted to influence the relevant blocks, providing a bounded attribution set substantially narrower than the full corpus. The log thereby enables compliance auditing: a content owner's inquiry whether their content was used receives a definitive answer, and a regulator's requirement that restricted content was not deeply integrated is satisfied by the depth profile records.

Training-Level Memorization Detection

The provenance log and depth-selective aggregation records enable a memorization detection mechanism that determines exactly where and how deeply similar training content was integrated into the parameters. When model output at inference time is flagged as exhibiting high similarity to a known training artifact, the detection module initiates a reverse provenance query, identifies the training examples corresponding to the artifact, and retrieves their depth-aggregation profiles, per-layer contribution weights, and governance records. From these records the system determines which layer blocks the similar content was permitted to influence, the magnitude of the gradient that reached each block, the entropy band and policy scope under which the content was admitted, and whether it was admitted under a suppressed or full-depth profile.

The module produces a structured assessment classifying the similarity as one of three categories. Shallow memorization indicates the content was trained with a suppressed profile confining its influence to shallower layers, a consequence of lexical or local structural similarity rather than deep conceptual encoding, and the expected outcome when rights-restricted content is properly governed. Deep memorization indicates the content's influence extends into deeper layers, which may reflect policy-compliant integration of freely licensed content or a governance failure in which depth-restricted content was inadvertently trained at full depth. Absent memorization indicates the log contains no record of the content, so the similarity is treated as coincidental. The assessment is reported to the inference-time governance substrate, which may permit the output with an attribution annotation, suppress it and generate a governance alert, or apply standard admissibility evaluation accordingly.

Integration with Inference-Time Governance

The training provenance records are made available to the inference-time semantic execution substrate, closing the governance loop between training time and inference time. When the inference substrate evaluates a candidate transition for admissibility, it may query the training provenance log to determine the training-time governance profile of the knowledge that grounds the transition, identifying the content whose depth profiles encompass the layer blocks most active during generation. If that content was admitted under a policy that has since expired, the substrate may apply heightened scrutiny or reject the transition. If it was admitted under a policy revoked by the content owner, the substrate may reject the transition and generate a governance alert. If it was admitted under a suppressed profile indicating rights restriction, the substrate may require an attribution annotation.

The training-time governance profile does not replace the inference-time evaluation; it enriches it with information about the provenance and governance of the knowledge the model is drawing upon. A model trained with depth-selective governance and queried through inference-time governance produces outputs that are doubly governed: the knowledge was governed at the point of integration, and the expression of that knowledge is governed at the point of emission. This continuous governance chain extends from the training corpus through the model's parameters to its inference output and is traceable, auditable, and enforceable at every stage.

Disclosure Scope

This disclosure covers the training loop reconceived as a governed execution environment, in which the semantic execution substrate evaluates each training example as a proposed semantic mutation and modulates the gradient signal across model depth according to entropy-band-indexed training depth profiles. It covers the depth-selective aggregation mechanics realized through gated residual connections, attention-based depth selection, or layer-specific scaling factors at block-level granularity; policy-governed knowledge retention and suppression through suppressed and zero-weight depth profiles, contrasted with post-hoc unlearning by structural prevention; the append-only training provenance log with its forward and reverse queries; training-level memorization detection classifying similarity as shallow, deep, or absent; and the integration of training-time provenance into inference-time admissibility. These mechanisms are disclosed in the cognition filing (U.S. Application No. 19/647,395 and its international counterpart). The scope extends to embodiments using any specific admissibility predicate, any depth profile representation, and any provenance serialization, provided the substrate governs the depth of gradient integration and produces a traceable training provenance record.