Mechanism
The disclosed mechanism extends the semantic execution substrate from inference time into the training loop itself, so that each training iteration is treated as a proposed semantic mutation to the model's accumulated knowledge state. In conventional training the loop is an ungoverned optimization process: data is sampled, forward and backward passes compute gradients, and the optimizer applies parameter updates with no intermediate admissibility evaluation. Every training example contributes to every layer's updates with equal structural authority. There is no mechanism to evaluate whether a given example should be permitted to influence the parameters, no mechanism to control the depth at which its contribution is integrated, and no mechanism to record which examples influenced which capabilities.
To govern this, the substrate is positioned 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 the substrate produces for the example that generated those gradients. 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. This mirrors the inference-time positioning, where the substrate intercepts candidate transitions between proposal and commitment.
Training Examples as Proposed Semantic Mutations
A semantic mutation, in the sense used throughout the filing, is a structured modification to a governed state that must be evaluated for admissibility before commitment. At inference time a mutation is a proposed transition in the output trajectory. At training time a mutation is a proposed modification to the model's learned representations, the statistical patterns encoded in the parameters that collectively constitute the model's knowledge. Treating each training example this way produces three consequences. First, non-training, the refusal to integrate an example into the parameters, is a valid computational result rather than an error condition: the substrate may determine an example is inadmissible and produce a training iteration in which the parameters are not updated, recording that non-training as a governed event with the example's identity and the reason for rejection.
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 derived from the platform's entropy extraction pipeline, a slope position indicating the content's place in the trust-slope hierarchy, a content provenance record identifying source, acquisition pathway, and chain of custody, and a policy scope identifying the governance constraints (licensing terms, usage restrictions, temporal validity bounds, exclusion mandates) that apply. Third, the training corpus itself becomes a governed collection of semantically annotated objects, each carrying its own governance profile, rather than an undifferentiated mass of data.
Graded Admissibility, Not Binary Admit or Reject
The substrate's evaluation is not limited to a binary admit or reject. It may render a graded determination that specifies not merely whether an example is admissible but how its contribution should be weighted, routed, and distributed across the model's depth. An example may be admitted for shallow integration but excluded from deep integration. It may be admitted with full depth access for one set of layers and suppressed access for another. It may be admitted with a reduced contribution magnitude that attenuates its influence relative to examples admitted at full magnitude. The per-example determination comprises an admit or reject decision together with a depth-aggregation profile, and admitted examples are routed through that profile to the aggregation mechanism for depth-selective parameter updates.
Entropy-Band-Indexed Depth Profiles
Each entropy band recognized by the platform is associated with a training depth profile that governs how content from that band is weighted across the model's layers during training. The band classification is derived from the semantic entropy of each example, the information-theoretic divergence of its semantic embedding distribution relative to the model's current representational state, such that low-entropy content (well represented in existing knowledge) receives shallow profiles and high-entropy content (introducing novel semantic structure) receives deep profiles. The depth profile is a structured object: a per-layer or per-block contribution weight vector. A weight of one permits the full gradient signal to reach a block, a weight of zero prevents any signal from reaching it, an intermediate weight attenuates the signal, and a weight greater than one amplifies it.
The association between bands and profiles is not fixed before training and held unchanged. A profile adaptation engine monitors the model's internal entropy distribution at defined checkpoints, evaluating the entropy of layer-wise activation distributions on a held-out set, and adjusts the profiles to keep the corpus's band structure aligned with the model's evolving internal structure. Early in training the profiles may be broad, with all bands contributing to all layers at approximately uniform weight. As representations stratify, with shallow layers encoding local patterns and deep layers encoding abstract relationships, the profiles narrow: low-entropy content is weighted toward shallow layers, high-entropy content toward deep layers. By restricting routine content to shallow integration the system preserves deep representational capacity for the complex content that requires multi-step abstraction to encode.
Depth-Selective Aggregation Mechanics
Depth-selective aggregation applies the depth profiles to the gradient signal during the backward pass, modulating each example's contribution at each layer transition according to the profile associated with its entropy band. This is distinct from layer-wise aggregation developed for federated learning, which assigns different weights to layers across multiple model instances being merged: the disclosed mechanism aggregates no models, it governs the depth at which a single example's gradient is integrated into a single model. The mechanism is realized through one or more of three complementary techniques. Gated residual connections augment each residual pathway with a gating coefficient derived from the profile, applied during the backward pass only so that forward-pass inference behavior is unaffected. Attention-based depth selection scales the gradient reaching the attention weights and value projections at each transformer layer by the per-layer profile weight. Layer-specific scaling factors, an architecture-agnostic approach, multiply the gradient signal at each layer boundary by the profile weight before accumulation, requiring only the ability to intercept and scale the gradient.
The mechanism operates at block-level granularity rather than per individual layer: contiguous layers performing a coherent function are grouped into blocks, and the profile specifies per-block weights, balancing expressiveness against the cost of profile evaluation. In the gradient accumulation process, each example's gradient is first scaled by its per-block profile weight before being accumulated into the block's gradient buffer, so that per-example scaling occurs after per-example gradient computation and before batch-level accumulation. The mechanism does not alter the optimizer's update rule and is compatible with stochastic gradient descent, Adam, AdamW, and their variants: it alters the gradient signal the optimizer receives, not how the optimizer consumes it.
Policy-Governed Retention and Structural Prevention
The depth-selective mechanism is integrated with the platform's content governance so that the same policy objects governing agent behavior and inference-time admissibility govern depth and magnitude of integration during training. Content admitted under time-limited licensing is trained with a suppressed depth profile, one in which contribution weights for deeper layers are set to zero or near-zero, confining the example's influence to shallower layers so that future de-emphasis can be achieved through targeted shallow-layer adjustment rather than model-wide intervention. Content from the governed exclusion corpus is structurally prevented from any integration through a zero-weight profile that sets the contribution weight to zero at every layer, the training-time analog of the inference-time rejection determination, recorded in the provenance log as a governed event.
This is structural prevention, distinct from post-hoc unlearning. Post-hoc unlearning operates after training: it approximates the influence of content that should not have been learned and applies corrective updates intended to remove it, and it is inherently approximate because a single example's influence is diffused across millions or billions of parameters through the non-linear dynamics of optimization. The disclosed approach prevents the gradient from reaching the deep layers in the first place, so there is no need to unlearn what was never deeply learned. The prevention is exact rather than approximate: a zero weight at a given block means no gradient from the example reaches that block's parameters. Where multiple policies apply to one example the substrate resolves the profile by applying the most restrictive policy, and the resolution is recorded for audit. The result is a model whose internal knowledge structure reflects its governance constraints: freely licensed content encoded deeply and durably, restrictively licensed content encoded shallowly and separably, excluded content not encoded at all.
Provenance Log and Memorization Detection
The substrate records a comprehensive provenance trail for every training iteration, the training-time analog of the lineage field maintained for 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-block contribution weight that actually reached each block after any adaptation, the governance record identifying the policy objects that authorized admission and determined the profile, the content provenance record, and the admissibility determination (admitted, rejected, or admitted with modified profile) with the reason for any modification. The log is chronologically ordered and append-only, each entry timestamped, sequentially numbered, and annotated with epoch, iteration, and batch index, so entries cannot be retroactively modified, deleted, or reordered without detectable inconsistency. A forward query traces a given example through the profiles and weights that governed its integration, identifying which blocks it influenced and with what magnitude. A reverse query begins from an observed behavior and identifies the set of training content that was structurally permitted to influence the active blocks, a bounded attribution set substantially narrower than the full corpus, though the non-linear dynamics of optimization preclude exact attribution.
These records enable training-level memorization detection. When inference output is flagged as exhibiting high similarity to a known training artifact, a reverse query retrieves the corresponding examples' profiles, weights, and governance records, and the system classifies the similarity into one of three categories. Shallow memorization indicates the content was trained with a suppressed profile confining it to shallow layers, the expected outcome when rights-restricted content is properly governed. Deep memorization indicates the content's influence extends into deep layers, which may be policy-compliant for freely licensed content or may indicate a governance failure where depth-restricted content was inadvertently trained at full depth. Absent memorization indicates the log contains no record of the content, so the similarity is not a consequence of direct training on the artifact. The assessment is reported to the inference-time governance layer, which incorporates it into its admissibility determination.
Composition With the Rest of the Architecture
Training-time governance composes with inference-time governance through the shared provenance infrastructure, closing the loop between the two stages. When the inference substrate evaluates a candidate transition it may query the training provenance log for the governance profile of the knowledge grounding that transition: the policy objects that governed the content's admission, the profile applied, and any temporal validity constraints. If that content was admitted under a policy that has since expired the substrate may apply heightened scrutiny, if under a policy revoked by the owner it may reject the transition and raise an alert, and if under a suppressed profile it may require attribution. The result is doubly governed output: knowledge governed at the point of integration and governed again at the point of emission, a continuous chain from corpus through parameters to output. The same depth-selective mechanism is applied in further embodiments to curriculum-integrated depth scheduling, where the curriculum engine controls the temporal sequencing of content while the depth profiles control its spatial integration, to affect-modulated training depth, to per-content differential privacy via gradient routing, to real-time learning from individual interactions through a parameter-efficient adaptation layer, and to governed skill adapters discovered and composed through the adaptive index.
Disclosure Scope
The treatment of each training example as a proposed semantic mutation evaluated by the semantic execution substrate at the boundary between loss computation and gradient application, the graded admissibility determination comprising an admit or reject decision and a depth-aggregation profile, the entropy-band-indexed depth profiles and their adaptation by the profile adaptation engine, the depth-selective aggregation mechanics realized through gated residual connections, attention-based depth selection, or layer-specific scaling factors, the policy-governed suppressed and zero-weight profiles that effect structural prevention as distinct from post-hoc unlearning, the append-only training provenance log with its forward and reverse queries, and the training-level memorization detection feeding inference-time governance, are 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 depth-selective routing is realized over different model architectures using gradient-based optimization, and does not cover the optimization algorithms used to compute gradients or the underlying cryptographic sealing primitive, which are substitutable.
