Training Governance for Legal AI: Encoding Precedent Hierarchy Into the Model

The problem: a legal corpus is not a flat corpus

Legal authority is hierarchical and time-indexed in ways that ordinary text is not. A United States Supreme Court holding binds every court below it; an intermediate appellate decision binds only within its circuit or district; a decision from a sister jurisdiction is at most persuasive; a dissent states the law that did not prevail; and an overruled or vacated opinion states law that is no longer good. A statute superseded by amendment is text that should not be relied on. None of this structure is legible to a training pipeline that treats each opinion as a token stream weighted by how often it appears.

The consequences are not hypothetical. Generative legal tools have produced briefs citing cases that do not exist, and the resulting sanction orders have reset what courts and bar regulators expect from systems used in practice. The professional and regulatory backdrop is dense: confidentiality and supervision duties under the rules of professional conduct, candor-to-the-tribunal obligations, verification expectations attaching to filed papers, state-bar formal opinions on generative tools, discovery and proportionality rules, and the designation of administration-of-justice systems as high-risk under emerging AI regulation. A model that cannot represent the difference between binding and persuasive authority cannot satisfy any of these regimes, and the firm deploying it carries the liability.

The market has responded mostly at the edges: retrieval filters that exclude bad citations after generation, prompt scaffolding that asks the model to "check" authority, and human review of every output. These are mitigations layered on top of a model whose parameters already encode the wrong thing. The defect is upstream, in training. Training Governance addresses it there.

What the disclosed technology does

Training Governance, disclosed in Chapter 11 of United States Patent Application 19/647,395, is a depth-selective training architecture. Rather than admitting every example into every layer of the model at uniform weight, it evaluates each training example through a semantic substrate, assigns it a structured training depth profile, and routes its gradient contribution to specific layers of the network according to that profile. The depth profile is a per-layer (or, for deep networks, per-block) vector of weights that determines how strongly a given example is allowed to shape shallow, middle, and deep representations.

Three primitives from the specification carry the legal application:

Entropy-banded depth routing. The substrate evaluates each example's semantic entropy and assigns a depth profile accordingly. Low-entropy content, which reinforces patterns the model already holds, is weighted toward shallow layers that encode surface structure. High-entropy content, which introduces novel semantic relationships, is weighted toward deep layers that encode abstract relationships. The profiles can adapt during training: the substrate monitors the entropy of the model's own layer representations and adjusts routing to keep each depth band aligned with the kind of content it is meant to encode.

Depth-selective gradient aggregation. Routing is enforced during the backward pass through one or more of three complementary techniques disclosed in the spec: gated residual connections that attenuate or amplify gradient flow per layer according to the depth-profile weight; attention-based depth selection that modulates per-layer attention weighting by the depth-profile weight without altering the forward pass; and layer-specific scaling factors that multiply the computed gradient at each layer or block by the profile weight before accumulation. A zero-weight depth profile sets the contribution weight to zero at every layer, so an inadmissible example is evaluated, found inadmissible, and structurally excluded from influencing any parameter. This is the training-time analog of the inference-time admissibility gate's rejection determination.

Provenance-traceable training with reverse provenance. Every training iteration is recorded in an append-only, chronologically ordered, sequentially numbered and timestamped training provenance log that captures, for each example, a content provenance record (source, acquisition pathway, chain of custody, semantic metadata), the admissibility determination, and the depth profile applied. The log can be periodically sealed using the cryptographic sealing infrastructure cross-referenced from the Governance disclosure, producing tamper-evident checkpoints that support third-party verification. The log supports reverse provenance queries: given an observed model behavior, the query traces back through the log to identify the training content whose depth profiles encompassed the layer blocks active during that behavior. The same provenance infrastructure underpins memorization detection and inference-time governance.

Provenance and admissibility, not corpus frequency, govern which content trains which layers. That single inversion is what makes the architecture fit for law.

Applying it to legal authority

The legal application maps the authority hierarchy onto the depth profile. The mapping is a deployment-configured governance policy, not new model machinery; it is built entirely from the disclosed primitives.

• Binding precedent in a forum is admitted with a depth profile that reaches the model's deep layers, so its holdings shape the abstract legal representations the model reasons from. Authority is expressed as routing depth.
• Persuasive authority (sister-jurisdiction decisions, secondary sources, treatise commentary) is admitted with a shallower profile. It contributes surface familiarity and vocabulary without being permitted to drive the model's deep representation of what the law is in the governed forum.
• Dissents and concurrences are admitted with provenance metadata marking them as non-controlling, and routed so the model can recognize and reproduce their reasoning when asked while not encoding them at the depth reserved for the majority holding.
• Overruled, vacated, or superseded authority is assigned a suppressed or zero-weight depth profile. A suppressed profile zeroes the deeper-layer weights so the content cannot influence the model's operative representation of current law; a zero-weight profile excludes it from every parameter while still recording the evaluation in the provenance log. Either way the determination is deterministic and audited.
• Jurisdiction- and practice-area-scoped models follow from per-deployment policy: the same corpus yields different depth profiles under a federal-tax policy than under a state-family-law policy, because admissibility and depth are resolved against the governing policy object rather than fixed in the data.

Because routing is enforced in the backward pass and recorded as it happens, the authority hierarchy is encoded in the parameters themselves rather than reconstructed by a downstream filter. The model does not have to be reminded at inference time that a dissent is a dissent; the gradient that would have made it authoritative was attenuated at training time.

Why provenance is the defensible part

The professional-responsibility exposure around legal AI is ultimately evidentiary: a regulator, a court, or an adversary asks how the model came to assert what it asserted. The append-only, cryptographically sealable training provenance log answers that question in the form courts and auditors accept. Reverse provenance can take an observed assertion and identify the training content and depth profiles behind it. When a model declines to treat an overruled case as good law, the log shows the suppressed depth profile and the admissibility determination that produced that result, with a sealed, tamper-evident checkpoint a third party can verify. This converts "trust the vendor" into an auditable record, which is the posture the governance and high-risk-system regimes are converging on.

Embodiments and deployment options

The application is not a single instance. The disclosed primitives support a range of grounded embodiments:

• Per-forum models that hard-encode a single jurisdiction's binding hierarchy, versus multi-jurisdiction models that carry several hierarchies distinguished by provenance metadata and resolved by policy at inference.
• Continuous re-grounding, in which a court overruling a precedent triggers a fine-tuning pass that reassigns the affected authority a suppressed or zero-weight depth profile, with the change recorded in the provenance log so the model's law tracks the law.
• Routing enforced through any of the three disclosed aggregation techniques (gated residual connections, attention-based depth selection, or layer-specific scaling), chosen to fit the base architecture, including the per-block variant for very deep networks where per-layer profiles would be unwieldy.
• Confined training of sensitive material, where privileged or client-confidential content is admitted only at shallow, non-memorizing depth or excluded by zero-weight profile, using the platform's memorization detection and differential-privacy-by-confinement to keep deep representations from absorbing text that should never be reproduced verbatim.
• Audit-first deployments for regulated or high-risk settings, where the sealed provenance log and reverse-provenance queries are exposed as a compliance interface for examiners and opposing counsel.

A skilled implementer building on the disclosed depth-profile router, depth-selective aggregation, and provenance log can construct any of these without inventing new mechanism. The legal-authority mapping is policy configuration over technology that already exists in the specification.

Disclosure Scope

This article describes an application of the depth-selective training governance disclosed in United States Patent Application 19/647,395. The market framing, the legal-authority mapping, and the deployment scenarios are illustrative applications. The underlying technical mechanisms (depth-profile routing, entropy-banded depth assignment, depth-selective gradient aggregation via gated residual connections, attention-based depth selection, and layer-specific scaling, suppressed and zero-weight depth profiles, memorization detection, the append-only cryptographically sealable training provenance log, and reverse provenance queries) are disclosed in that application. No training metrics, dataset sizes, or benchmark numbers are claimed.