Policy-Governed Knowledge Retention and Suppression

Mechanism

Knowledge retention in this disclosure is not a separate retention engine layered over the optimizer. It is a consequence of the depth-selective aggregation mechanism that governs training. Each training example is treated as a proposed semantic mutation to the model's knowledge state, and the semantic execution substrate evaluates that example before its gradient signal is permitted to affect the model's parameters. 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 according to the admissibility determination the substrate produces. Retention and suppression are therefore decided by where, and at what magnitude, a training example's gradient is allowed to reach the model's layers.

The substrate produces, for each example, a depth profile: a per-layer or per-block contribution weight vector. A weight of one permits the full gradient signal to reach that block, a weight of zero prevents any gradient from reaching it, and an intermediate weight attenuates the signal by the specified factor. The depth profile is the structural lever for retention. Content that should be encoded deeply and durably is admitted with a profile that permits deep integration. Content that may need to be de-emphasized or removed later is admitted with a suppressed profile that confines it to shallow layers. Content that is inadmissible is admitted with a zero-weight profile and is not encoded at all. Retention is thus governed at the point of integration rather than recovered after the fact.

Suppressed and Zero-Weight Depth Profiles

Content admitted to the training corpus 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 example's influence to the shallower layers. The rationale is structural separability: content whose license may expire, whose owner may revoke permission, or that may later be identified as inaccurate should not be deeply encoded, because removal from deep parameters would otherwise require invasive procedures such as full retraining or approximate unlearning. By confining time-limited content to shallow layers, the system makes the content's influence structurally separable from the model's deep knowledge, so that de-emphasis can be achieved through targeted shallow-layer adjustment rather than model-wide intervention.

Content from the governed exclusion corpus, identified by the platform's governance infrastructure as inadmissible for training, is structurally prevented from deep integration through a zero-weight depth profile that sets the contribution weight to zero at every layer, preventing the example from influencing any of the model's parameters. The zero-weight profile is the training-time analog of the rejection determination in the inference-time admissibility gate: the content is evaluated, found inadmissible, and structurally excluded from the model's knowledge. The exclusion is recorded in the training provenance log as a governed event, including the identity of the excluded content and the policy that mandated exclusion. Non-training, the refusal to integrate an example, is a valid computational result rather than an error condition.

Structural Prevention Rather Than Post-Hoc Unlearning

The distinction between this disclosure and post-hoc unlearning is structural. Post-hoc unlearning operates after the model has been trained: it identifies content that should not have been learned, approximates that content's influence on the parameters, and applies corrective updates intended to remove or attenuate it. That approach is inherently approximate, because the influence of any single training example on a deep network's parameters is diffused across millions or billions of parameters through the non-linear dynamics of gradient-based optimization, making it impossible to precisely identify and reverse the parameter changes attributable to a specific example.

The present disclosure does not rely on post-hoc unlearning. It implements structural prevention: 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 exact rather than approximate, because the per-block weights are applied to the gradient signal at each block boundary and a zero weight at a given block means no gradient from the example reaches that block's parameters. The prevention is deterministic, auditable, and reversible by changing the depth profile, whereas post-hoc unlearning is stochastic, approximate, and irreversible.

Policy-Governed Retention

The policy objects that govern depth profiles for content classes are the same policy objects that govern agent behavior and inference-time admissibility throughout the platform. A policy object that specifies rights-grade constraints for a class of content, for example a policy stating that content from a particular creator is admitted under a license that expires on a defined date, is consulted by the semantic execution substrate during training to determine the depth profile for examples drawn from that class. When the policy is active, the substrate applies the suppressed depth profile the policy specifies. When the policy expires, the substrate applies a zero-weight depth profile, preventing further integration. When the policy is revoked by the content owner, the substrate applies a zero-weight depth profile and records the revocation event in the training provenance log.

When multiple policies apply to a single training example, for example an example simultaneously subject to a creator's licensing policy, a domain-specific regulatory policy, and a platform-wide governance policy, the substrate resolves the applicable depth profile by applying the most restrictive policy. If the creator's policy permits deep integration but the regulatory policy restricts integration to shallow layers, the regulatory restriction prevails. The hierarchical resolution logic is deterministic and is recorded in the training provenance log, so a post-training audit can determine which policy governed the depth profile for each example. The result is a model whose internal knowledge structure reflects the governance constraints under which it was trained: freely licensed content is encoded deeply and durably, content admitted under restrictive licenses is encoded shallowly and separably, and excluded content is not encoded at all.

Provenance and Memorization Detection

The semantic execution substrate records a comprehensive provenance trail for every training iteration, producing a training provenance log that enables post-training analysis to reconstruct which content influenced which model capabilities at which depths. For each batch or example the log records, among other elements, the entropy band classification, the slope position, the depth aggregation profile that was applied, the per-block contribution weight that actually reached each block, the governance record identifying the policy objects that authorized admission and determined the depth profile, the content provenance record, and the admissibility determination. The log is chronologically ordered and append-only, so entries cannot be retroactively modified, deleted, or reordered without producing detectable inconsistencies in the sequential numbering and timestamps.

The provenance log supports a training-level memorization detection mechanism. When model output at inference time is flagged as exhibiting high similarity to a known training artifact, the memorization detection module initiates a reverse provenance query that retrieves the corresponding examples' depth profiles, contribution weights, and governance records, and classifies the similarity into one of three categories. Shallow memorization indicates the content was trained with a suppressed profile confining its influence to shallower layers, the expected outcome when time-limited or rights-restricted content is properly governed. Deep memorization indicates the content's influence extends into the deeper layers, which may be policy-compliant for freely licensed content or may indicate 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 not a consequence of direct training on the artifact.

Separability as an Architectural Guarantee

Because retention is expressed through depth profiles, the durability and separability of each knowledge component can be set per content class and per domain. Embodied robotic agents trained with depth-selective governance segregate safety-critical knowledge from preference-based knowledge through depth-profile differentiation. Safety-critical motor control knowledge, including obstacle avoidance, emergency stop protocols, human proximity detection, and force-limiting behaviors, is trained with deep, protected profiles so that it is durably encoded in the deepest layers, where it is least susceptible to disruption by subsequent training or fine-tuning. Preference-based knowledge, including user-specific motion preferences and aesthetic movement patterns, is trained with shallow profiles that confine it to layers that can be updated or retrained without affecting the safety-critical deep representations.

The same separability supports therapeutic agents, where specific clinical protocols that vary across jurisdictions or are subject to ongoing regulatory revision are encoded with suppressed profiles confining them to intermediate layers so they can be updated or replaced without full model retraining, while patient-specific content, if used at all, is encoded with maximally suppressed profiles under time-limited policy governance that ensures automatic exclusion upon policy expiration. In each case the training provenance log lets an auditor verify that the content was governed in accordance with the applicable requirements.

Prior-Art Distinction

Conventional training is an ungoverned optimization process: every training example contributes to every layer's parameter updates with equal structural authority, with no mechanism to evaluate whether an 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. Machine-unlearning techniques attempt to address removal after training, but operate approximately because example influence is diffused across the parameters by the non-linear dynamics of optimization. The retention mechanism disclosed here differs by preventing deep integration at training time rather than reversing it afterward, making removal of suppressed content unnecessary because the content was never deeply encoded.

Differential privacy for machine learning typically adds noise uniformly to all gradient signals, calibrated to the worst-case privacy requirement across the entire corpus and degrading accuracy for content that does not require protection. The depth-selective approach instead routes privacy-sensitive content's gradients primarily to shallow layers, where representations are generic and inherently less memorizable, while suppressing contribution to deep layers, so the guarantee is structural rather than statistical: the model cannot memorize what it was not permitted to encode in memorizable layers, and non-sensitive content can still be trained at full depth without the accuracy-privacy tradeoff inherent in global noise injection.

Disclosure Scope

Policy-governed knowledge retention and suppression, comprising the suppressed depth profile that confines time-limited content to shallow layers, the zero-weight depth profile that structurally excludes inadmissible content, the structural prevention of deep integration as a deterministic and reversible alternative to approximate post-hoc unlearning, the hierarchical most-restrictive policy resolution, the training provenance log and its reverse-query-driven classification of shallow, deep, or absent memorization, and the use of depth-profile differentiation to segregate durable safety-critical knowledge from separable preference-based or rights-restricted knowledge, is disclosed in the cognition filing (U.S. Application No. 19/647,395 and its international counterpart) at Sections 11.5 through 11.7, with the underlying depth-selective aggregation and provenance infrastructure at Sections 11.2 through 11.6. This article describes that disclosed mechanism. The scope extends to embodiments across single-model, fine-tuning, on-device adaptation-layer, and modular skill-adapter architectures, provided retention is governed by depth-selective routing of the gradient signal under policy control with provenance recording.