The problem a searcher is trying to solve
If you run a training pipeline, "did this specific document influence the model, and if so how deeply" is a question you usually cannot answer after the fact. Alignment training shapes behavior well, but it does not, by itself, produce a per-example record of which content was permitted to reach which layers. This article compares Anthropic's constitutional training approach with the depth-selective training governance disclosed in United States Patent Application 19/647,395, and scopes the comparison to that provenance-and-depth axis rather than to model quality or safety outcomes.
What Anthropic's constitutional training does well
Anthropic has publicly documented Constitutional AI: a training method in which a written set of principles (the constitution) guides model behavior through supervised fine-tuning followed by reinforcement learning from AI feedback (RLAIF), including a self-critique-and-revision loop in which the model critiques and revises its own outputs against the stated principles. The approach reduces reliance on large volumes of human preference labels for harm avoidance, makes the governing principles legible and publicly discussable, and has produced the Claude model family that many practitioners treat as a reference point for well-behaved general-purpose assistants.
These are real strengths, and the comparison below is not a claim that constitutional training is deficient at what it targets. It targets model behavior, and it aligns behavior. The axis this article addresses is a different one: the structural relationship between a training example and the parameters it shapes.
The architectural axis: content-to-layer routing and its record
Constitutional training, like standard pre-training and RLHF, aggregates gradient signal across the model's depth. Principles are internalized as distributed statistical regularities. That is an efficient and effective way to shape behavior. It is not, by construction, a mechanism that decides per training example which layers that example is permitted to influence, nor one that emits a durable, queryable record of that decision. This is a general architectural property of gradient-based training pipelines, not a defect specific to any one lab.
United States Patent Application 19/647,395 discloses training governance (Chapter 11) that operates on exactly this axis. The core mechanisms disclosed are:
Training examples as governed semantic mutations. Each training example is evaluated by a semantic substrate that renders a graded admissibility determination: not merely admit or reject, but a depth-aggregation profile specifying how the example's contribution is weighted and routed across the model's layers.
Entropy-band-indexed depth profiles. Each entropy band recognized by the platform is associated with a training depth profile: a per-layer or per-block contribution-weight vector. Low-semantic-entropy content (well represented in the model's current state, measured by a KL- or Jensen-Shannon divergence) receives shallower profiles; high-entropy content introducing novel structure receives deeper profiles. A profile-adaptation engine adjusts these profiles at checkpoints as the model's internal entropy distribution evolves.
Depth-selective gradient routing. The per-layer weight vector modulates the gradient signal each example is permitted to contribute at each layer or residual block during backpropagation. A weight of one passes the full signal, zero blocks it, values between attenuate, and values above one amplify. The disclosure notes the modulated gradient buffer is consumed by the optimizer as an ordinary gradient buffer, so the mechanism integrates into existing pipelines.
Differential privacy by architectural confinement. As disclosed in Section 11.12, privacy-sensitive content is routed primarily to shallow, generic layers and suppressed from deep, memorizable layers. The guarantee is structural rather than statistical: the model cannot memorize what it was not permitted to encode in memorizable layers, avoiding the uniform accuracy-privacy tradeoff of global noise injection.
Append-only, sealable training-provenance log. Section 11.6 discloses a chronologically ordered, sequentially numbered, timestamped append-only log recording each example's provenance, depth profile, contribution weights, admissibility determination, and policy scope. The log is periodically sealed using the cryptographic sealing infrastructure of the cross-referenced governance filing, producing tamper-evident checkpoints for third-party verification.
Memorization detection and reverse provenance. Section 11.7 discloses that when an inference output is flagged as similar to a known training artifact, a reverse query against the provenance log retrieves the artifact's depth profile and contribution weights and classifies the similarity as shallow memorization (confined to shallow layers), deep memorization (integrated into deep layers), or absent (no training record). The classification feeds the inference-time governance substrate.
Governed fine-tuning provenance. Section 11.13 discloses recording fine-tuning admissibility and depth profiles as a provenance record structurally distinct from pre-training, so that a challenged output can be attributed to parameter regions shaped by pre-training versus fine-tuning content, supporting liability allocation in regulated deployments.
Where the two approaches differ, stated fairly
Constitutional training answers "does the model behave according to these principles." Depth-selective training governance answers a different, complementary question: "which content was permitted to influence which layers, under what policy, and can that be proven after the fact." A constitution recited in a system prompt binds behavior at runtime through the trained weights; it does not, by itself, produce a cryptographically sealed, per-example, depth-resolved provenance trace from a runtime output back to the training events that shaped it. That trace is the specific contribution disclosed in 19/647,395. Neither approach subsumes the other, and a governed pipeline could in principle carry constitutional-style alignment objectives while adding the routing and provenance machinery above.
Enablement and embodiments
A skilled implementer can build the disclosed approach. Depth profiles are per-layer or per-block weight vectors; routing is applied by multiplying each example's per-layer gradient by its profile weight before the optimizer step, which the disclosure notes leaves the optimizer interface unchanged. Entropy bands are computed from the divergence between an example's semantic embedding distribution and the model's current representational state. The provenance log is an append-only, sequentially numbered, timestamped store with periodic cryptographic sealing.
Disclosed embodiments and variations include: profiles that adapt at checkpoints via the profile-adaptation engine versus fixed profiles; per-layer versus per-block granularity; attenuating, blocking, and amplifying weights; differential privacy by confinement for privacy-sensitive content; curriculum-integrated depth scheduling ordered by governance properties rather than difficulty; content anchoring by structural entropy so provenance-incomplete content can be restricted to shallow layers; the three-way memorization classification driving permit-with-attribution, suppress-and-alert, or standard evaluation at inference; and governed fine-tuning provenance for pre-training-versus-fine-tuning liability attribution. These mechanisms apply across general-purpose and applied (for example medical or robotics) training contexts.
Disclosure Scope
The invention described here is disclosed in United States Patent Application 19/647,395 (Training Governance). All statements about what the platform does, its mechanisms, and its guarantees are grounded in that filing. References to Anthropic, Constitutional AI, RLAIF, and the Claude model family are external context describing a third party's publicly documented approach; they are provided for architectural comparison and are not claims of the filing, and they should not be read as characterizing that party's training-data sources, licensing, or legal exposure. Named products and companies are the property of their respective owners. This article is a dated public disclosure tied to the referenced application.
