The Consistency Problem in Automated Legal Advice
Legal advice is a coherence-bound practice. A position taken in one memorandum constrains the positions a lawyer can responsibly take in related matters; advice must track binding authority; and a representation made to a tribunal carries a certification that the signer has formed a competent, non-frivolous view. A stateless inference engine satisfies none of these constraints structurally. It produces each output from the prompt in front of it, retains no persistent record of the positions it has already committed to, and offers no internal account of whether a new conclusion is consistent with cited authority or with its own prior reasoning. The lawyer is left to reconstruct that account by hand, after the fact, under a certification duty that does not distinguish between human and machine error.
The Integrity and Coherence step disclosed in United States Patent Application 19/647,395 (Chapter 3) addresses this by treating behavioral consistency as a computable field carried inside the agent object itself, evaluated continuously, recorded in lineage, and capable of triggering self-correction before an inconsistent output is delivered. The remainder of this article maps each disclosed primitive onto a concrete legal-advisory deployment.
Integrity as a Three-Domain Field
The disclosure structures the integrity field as three independently tracked domains: personal, interpersonal, and global. In a legal-advisory agent these map cleanly onto the duties the practice imposes.
Personal integrity encodes the agent's self-referential alignment, the degree to which its actions are consistent with its own declared values and self-imposed constraints, evaluated by comparing its behavioral record in lineage against a declared value set held in the policy reference field. For a legal agent, the declared value set encodes commitments such as citing only authority it has verified, flagging adverse controlling precedent, and not reversing a previously stated legal position without recording the basis for the change. When the agent drafts an analysis that omits a controlling case it earlier relied on, the personal integrity score decreases in proportion to the significance of the inconsistency.
Interpersonal integrity encodes relational consistency, evaluated against the relational norms in the policy configuration, the expectations encoded in active delegation contracts, and prior interactions with the same entities. The disclosure's own example is an agent delegated a task with a specified confidentiality scope that discloses information outside that scope, which decreases the interpersonal score. This is the Rule 1.6 surface directly: a matter delegated under a confidentiality scope and an active engagement context becomes a relational commitment the field tracks.
Global integrity encodes alignment with broader systemic and ethical norms, evaluated against system-level policy constraints and the downstream consequences projected by the semantic harm resolver. For a legal agent, the global domain carries the duty-of-candor and unauthorized-practice constraints that transcend any single client relationship.
The three domains are tracked independently, so an agent may show high personal integrity yet low interpersonal integrity. Where evaluation requires a single figure, the disclosure combines the domains into a composite integrity score using deterministic, policy-specified domain weights. A policy governing privileged client communications can weight the interpersonal domain heavily; a policy governing representations to a tribunal can weight the global domain heavily. The weighting is fixed by policy, not negotiated by the agent, which is the property that makes the resulting record defensible.
Detecting an Inconsistent Position Before It Is Delivered
Consistency detection rests on the deviation function disclosed in the same chapter. The function is the deterministic composite D = (N - T) / (E x S): N is the agent's current need vector, T its ethical threshold, E the empathy weighting, and S the self-esteem score. The numerator is deviation pressure, the denominator deviation resistance. The function is evaluated continuously as part of the cognitive cycle, not as a periodic audit, so the conditions for an inconsistent output can be observed accumulating before the output is committed.
In a legal-advisory deployment, deviation pressure rises when a requested output would require the agent to depart from a position recorded in its lineage or from controlling authority it has registered, for example when an instruction pushes toward an aggressive reading that contradicts an earlier conservative opinion in a related matter. Because D is read at each decision point, the agent can hold the draft, surface the conflict, and route it for resolution rather than emit two irreconcilable opinions. When D crosses the activation threshold, the agent enters the Deviation-Activated State, in which mutations are scoped, the integrity field is updated, and the event is recorded before the state exits.
The Coherence Trifecta and Graded Self-Correction
When a deviation is detected, the coherence trifecta, the three-phase corrective loop of empathy, integrity, and self-esteem disclosed in the chapter, drives correction. A deviation event triggers empathy registration (projection of the consequence of the inconsistent advice), integrity recording (the deviation is written to the integrity field), and self-esteem-driven corrective pressure that produces a restorative mutation feeding back to reduce future deviation. Applied to legal advice, this means a detected contradiction does not merely raise a flag: it generates a candidate correction, such as a revised analysis that reconciles the new conclusion with the prior position or expressly distinguishes it, and records the reconciliation.
The disclosure provides coping intercepts on this loop at early, mid, and late points, corresponding to the empathy, integrity, and restoration phases, each leading to a stable outcome. This gives a deployment several configurable intervention points: an early intercept can hold an output the moment consequence projection flags it; a late intercept can let the restoration phase generate a corrected draft. Correction is graded, not absolute. Rather than a binary block, the agent's response scales with the magnitude of the inconsistency, which matches legal practice, where a minor citation gap and a position reversal call for proportionate responses.
Consistency Across Time: Trust Slope and Forecasting
Consistency in legal work is longitudinal. The disclosure applies trust-slope validation to the agent's own behavioral trajectory, producing an integrity trust score that measures the consistency between declared norms and observed behavior over time. A legal agent that has reliably flagged adverse authority and held consistent positions accumulates a high integrity trust score; an erratic one does not, and in multi-agent settings its contributions are weighted accordingly. The moral trajectory forecasting module projects the integrity trajectory across the three domains over future horizons and classifies it into archetypes such as a stabilization arc or a containment arc, generating containment recommendations when the projection indicates drift. For a supervising attorney, this is an early signal that an agent's positions are diverging before any single output crosses a line.
Deployment Options and Variations
The mechanism admits several enabling configurations. Domain weights can be tuned per practice area: litigation work weighting the global candor domain, transactional work weighting the interpersonal commitment domain. The declared value set in the policy reference field can encode firm-specific or jurisdiction-specific norms, so the same primitive serves a public-defender workflow and a regulatory-compliance workflow under different value sets. Coping intercept placement can be set conservatively (early intercept, hold-and-escalate) for high-stakes filings or permissively (late intercept, auto-reconcile-and-record) for internal research drafts. Activation thresholds and the pre-deviation margin are policy-specified, letting a firm calibrate how much pressure toward an inconsistent position is tolerated before the agent suspends and surfaces the conflict. Across all of these, every integrity field change, deviation event, scoped mutation, and restoration is recorded in the agent's lineage, producing the auditable trail a Rule 11 or Opinion 512 certification requires. The record is generated as the agent works, not assembled afterward, which is the structural advantage over a guardrail applied after inference.
None of this requires the agent to be a regulated actor or to render legal judgments autonomously. It is a structural support for a supervising lawyer: it keeps the machine's contribution internally consistent and self-documenting, leaving competent professional judgment, and the certification, with the human.
Disclosure Scope
The application of integrity and coherence to legal advisory agents described here is an enabling implementation of the Integrity and Coherence inventive step disclosed in United States Patent Application 19/647,395, including the three-domain integrity field (personal, interpersonal, global) and its deterministic policy-weighted composite, the deviation function D = (N - T) / (E x S) evaluated continuously across the cognitive cycle, the Deviation-Activated State and its scoped mutations, the coherence trifecta of empathy, integrity, and self-esteem with coping intercepts at the early, mid, and late phases, graded rather than absolute correction, trust-slope validation of the agent's own behavioral trajectory, moral trajectory forecasting, and lineage recording of all integrity state changes. The legal-domain framing, the regulatory mapping, and the deployment configurations are application context layered on that disclosed technology and are not claimed here as separate inventions. The scope is not limited to any particular practice area, value set, domain weighting, or intercept configuration; the structural commitment is the use of a continuously evaluated, lineage-recorded integrity field to keep an advisory agent consistent with cited authority and with its own prior positions before an output is delivered.
