top of page

Computable Due Process Proposal Sets Out Reasoning Tests for Judicial AI

26 minutes ago
11 min read
Illustrative photograph. Photo: Christian Wasserfallen / Pexels.

The Computable Due Process proposal, published November 11, sets out five conditions for AI-assisted judicial reasoning, while reserving final assessment for judges. Its central concern is the gap between a convincing legal document and a justified decision. An AI system can produce the former without establishing the latter.

The paper appears in Frontiers in Artificial Intelligence as a Hypothesis and Theory article. It introduces a conceptual specification, meaning a statement of properties an acceptable output should possess. It does not announce a deployed judicial model, an accuracy breakthrough, or a newly binding court standard.

For court software developers and institutional buyers, that distinction changes the discussion. The competing approaches are fluent document generation and inspectable legal justification. Existing guidance from England and Wales already emphasizes judicial responsibility. This proposal asks what reviewers should be able to inspect before accepting an AI-assisted draft.

The Computable Due Process Proposal Formalizes an Existing Judicial Duty

The paper proposes translating the duty to give reasons into properties that software and human reviewers can examine.

Roberto Leonardo Cruz-Núñez, Hugo Armando Aguilar-Brito, Andherson J. Aguirre-Lanegra, and Nuccia Seminario-Hurtado developed the framework. Their affiliations span Universidad Autónoma de Chiapas in Mexico and Universidad Tecnológica del Perú. Their computable due process paper starts from Peruvian constitutional case law and examines comparative legal materials.

The authors identify five conditions: an argument must exist, its inference must be valid, and its premises must have justification. It must also meet the applicable evidentiary standard and address the claims within the dispute. These conditions concern the relationship between facts, rules, objections, and conclusions.

That focus matters because legal drafting and legal justification are different tasks. A document can contain headings, citations, and an orderly narrative while skipping the step connecting evidence to its conclusion. Professional presentation does not supply the missing connection.

The publication therefore adds a proposed way to describe acceptable reasoning. It does not give courts a finished application that automatically checks every condition. The authors explicitly report that they conducted no empirical experiments and implemented no system.

Readers should also distinguish a proposed specification from the legal duties it interprets. Courts already operate under requirements to justify decisions within their jurisdictions. The researchers argue that those requirements can inform computational design. Publishing that argument does not independently change procedural law.

The immediate news is narrower than a headline about establishing judicial AI standards might suggest. Researchers have articulated a framework that developers can attempt to implement and test. Whether their reconstruction captures the relevant legal requirements remains open to scholarly and institutional evaluation.

This also explains why the work belongs in the judicial AI discussion despite lacking a product launch. Before measuring a system, developers need to specify what successful performance means. A polished draft and a reviewable argument represent different definitions of success.

For technology buyers, the proposal creates a useful distinction between capability demonstrations and acceptance criteria. A demonstration shows what an application can produce under selected conditions. Acceptance criteria determine what an institution requires across the situations where it intends to use that application.

That distinction becomes consequential when a draft reaches someone authorized to sign it. The document’s appearance can encourage acceptance before its underlying reasoning receives equivalent scrutiny. The framework directs attention to that handoff between generated text and institutional authority.

Judicial Responsibility Creates Pressure on Software Design

Human review becomes meaningful only when the reviewer can examine the material supporting the proposed conclusion.

The most direct pressure falls on developers offering judicial drafting assistance and institutions considering those products. Their challenge extends beyond making text editable. They must consider whether a reviewer can identify an unsupported premise, a missing response, or an unexplained inference.

England and Wales provides an institutional reference point. Its October 2025 judicial AI guidance reinforces judicial office holders’ responsibility for material produced in their names. It also places AI use within the obligation to protect judicial integrity and the rule of law.

The implication for interface design is straightforward, although it remains an analytical inference. A reviewer needs access to the evidence and authorities behind a draft. Requiring responsibility while hiding those materials would make substantive verification harder.

Consider a hypothetical application that generates a proposed ruling with citations beside each paragraph. That presentation helps users locate legal authorities. It does not necessarily show which evidence supports a factual finding or why the cited authority governs the particular dispute.

A different interface could expose those relationships directly. Selecting a conclusion might reveal its factual premises, the rule applied, and the objections considered. This would give the reviewer identifiable points to challenge, rather than requiring reconstruction from polished prose.

The difference also affects procurement. Asking whether a product includes human oversight yields limited information. Buyers need to understand what that oversight involves, what information the reviewer receives, and whether rejected reasoning remains visible in the record.

Time savings require similar scrutiny. Faster drafting does not automatically mean faster verification. An application might produce a document quickly while leaving the reviewer to check every factual and legal connection manually.

That observation does not establish that AI assistance increases judicial workloads. The paper supplies no operational trial capable of resolving that question. It does establish a reason to measure drafting time and review effort separately in future evaluations.

The pressure extends to lawyers and litigants because they encounter the reasoning that survives review. An affected party needs to identify why a submission failed. A general statement that the court considered everything offers little help when the decisive objection remains unanswered.

For knowledge workers, the broader connection concerns source organization. A searchable knowledge base can make supporting documents easier to retrieve. That organizational function remains distinct from determining whether a legal conclusion follows from those documents.

The institutional question is therefore specific: what must software reveal so that responsibility can be exercised? Answering it requires collaboration between legal specialists and developers. Neither fluent output nor a review button resolves it alone.

Inspectable Arguments Change What Legal AI Must Show

The proposed approach assigns different jobs to language generation, explicit reasoning structures, and judicial assessment.

The authors advocate an inspectable symbolic layer, meaning a representation of identifiable premises, rules, and relationships between conclusions. They leave room for language models to retrieve information and prepare drafts. The proposed reasoning structure carries the burden of displaying justification.

This is a division of responsibilities within a system, rather than a claim that document generation has no value. Finding relevant material and producing readable language remain useful functions. The disputed issue is whether those functions also establish an acceptable argument.

A hypothetical benefits dispute illustrates the difference. An application might retrieve an eligibility rule, summarize an applicant’s records, and draft a denial. The draft could still omit an exception that changes how the rule applies.

In an inspectable representation, the exception would need an identifiable place. The reviewer could examine whether its factual conditions were considered and why it applied or failed. That makes the disputed choice easier to locate.

The same principle applies when a factual premise remains uncertain. A system should distinguish an allegation from a finding accepted for the decision. Otherwise, later reasoning can appear internally consistent while relying on a premise nobody properly established.

Another distinction concerns the strength of the evidence. A confidence score describes a model’s assessment under its own design. It does not automatically correspond to the evidentiary requirement governing a judicial decision.

An adequate review process must therefore identify which question the number answers. Does it concern document classification, retrieval relevance, factual support, or something else? Combining those meanings into a single score would obscure rather than clarify the reasoning.

Historical research shows why specialization alone cannot settle reliability. Varun Magesh and colleagues evaluated legal research systems from LexisNexis and Thomson Reuters. Their legal AI reliability publication reported false information in more than 17% of responses from the evaluated tools.

Those results concern the products and versions tested. They should not serve as current performance estimates for every legal AI application. Nor do they measure the unimplemented framework described in the November paper.

The comparison nevertheless identifies different evaluation targets. A research assistant must find and characterize authorities accurately. A judicial drafting assistant must also connect applicable authorities and established facts to a proposed conclusion.

Retrieval-augmented generation, which supplies retrieved documents to a language model, addresses access to source material. Its presence does not independently establish that the generated interpretation is legally justified. A source can be genuine while the proposition attributed to it is mistaken.

The same limitation applies to lengthy explanations. More text can expand an argument, but it can also repeat an unsupported assumption. Reviewability depends on exposing the decisive relationships, rather than increasing the amount of prose.

The Computable Due Process proposal consequently shifts the design conversation toward inspectable commitments. What facts does the system assume? Which rule warrants the conclusion? Where does it address an objection capable of changing the result?

Those questions are useful precisely because they allow disagreement. A reviewer can reject a premise or interpretation without accepting the whole generated document. That gives substantive meaning to the ability to revise an AI-assisted draft.

A Formal Check Cannot Resolve Every Legal Judgment

The framework’s strongest limitation is explicit: satisfying its conditions does not guarantee a just decision.

The authors describe their conditions as necessary but insufficient. Choosing premises and weighing competing principles retain an evaluative component. A formal structure can reveal those choices without settling whether the choices deserve acceptance.

This creates a practical tradeoff for developers. Explicit representations offer inspectability, but someone must build and maintain them. A system cannot gain dependable legal coverage merely by replacing prose with boxes and arrows.

Consider a hypothetical conflict between a general rule and a narrower exception. Encoding a priority relationship makes the selected interpretation visible. The difficult question remains whether that priority is appropriate for the jurisdiction, facts, and authorities involved.

Maintenance introduces another challenge. A representation can remain internally consistent after its underlying legal source changes. Unless the system tracks those changes, an orderly argument may depend on an outdated assumption.

That is an engineering risk inferred from the proposed architecture, rather than a measured failure of an implemented product. The paper provides no deployment evidence establishing maintenance costs. Buyers should therefore treat operational feasibility as an unanswered question.

The framework also leaves room for disagreement over what counts as adequate support. Different reviewers can accept the same factual record while disagreeing about its significance. An application should expose that disagreement instead of disguising it as a technical certainty.

A further risk concerns the relationship between the displayed argument and the generated draft. If they diverge, the reasoning view could become another persuasive artifact. Evaluation should examine whether changes to supporting material alter the conclusions that depend on it.

That suggests a useful future test, rather than a reported result. Remove a decisive premise, introduce contrary authority, or reveal an unanswered objection. Then examine whether the application identifies the resulting weakness and presents it clearly to reviewers.

The paper’s broader claim about language models also requires careful handling. Evidence of hallucination, meaning false or unsupported generated content, establishes reliability problems in evaluated settings. It does not prove that every future architecture will fail every reasoning requirement.

Hybrid systems deserve assessment on their actual behavior. Combining a language model with explicit rules does not itself establish compliance. Conversely, using a language model somewhere in the workflow does not describe everything the application can verify.

Jurisdiction presents an additional boundary. The research builds from Peruvian doctrine and examines comparative materials. North American readers should not assume that its categories automatically reproduce every requirement in their own courts.

Transfer would require legal analysis of the relevant institution and proceeding. A shared interest in reasoned decisions does not erase differences in procedure, authority, or evidentiary obligations. Implementation would need to preserve those differences.

The skeptical question is therefore more demanding than whether formalization is possible. Can a maintained, usable implementation improve review across realistic disputes? The publication opens that question without supplying empirical evidence that resolves it.

Existing Court Rules Show Why Validation Needs Substance

The proposal enters an established governance discussion about control, accountability, and the ability to challenge AI-assisted work.

Spain offers a concrete legal comparison. Article 57 of its Royal Decree-Law 6/2023 addresses assisted judicial action. It provides that a generated draft does not independently become a judicial or procedural decision without validation by the competent authority.

The provision also requires that users can freely modify the draft. Its relevance here is the distinction between generating material and giving that material legal authority. The November research asks what substantive properties should support that validation.

These are related questions, but they operate at different levels. A rule can identify who holds responsibility and control. A proposed technical specification can describe information that would help that person exercise those responsibilities.

European ethical guidance supplies another reference point. The Council of Europe’s CEPEJ adopted its judicial AI charter in December 2018. Its principles include fundamental rights, nondiscrimination, quality and security, transparency, and user control.

That charter supports responsible judicial AI use while recognizing possible efficiency and quality benefits. It is not an endorsement of this particular paper. Its relevance lies in showing that accountability concerns preceded the latest generation of drafting systems.

The new proposal addresses a narrower design question within that wider discussion. What should a generated argument reveal so people can inspect it? Answering that question does not replace broader obligations concerning rights, security, or institutional independence.

For readers evaluating court technology, separating those layers prevents exaggerated claims. A system can offer detailed argument inspection while leaving other deployment questions unresolved. Likewise, an institution can adopt governance rules without possessing software that makes substantive review easy.

The distinction matters especially when discussing transparency. Publishing a general explanation of an application describes its intended operation. Showing the premises behind a particular conclusion allows someone to examine the reasoning in that individual matter.

Both forms of information can matter, but they serve different purposes. A person challenging a factual finding needs more than an architecture diagram. A procurement team assessing access controls needs information beyond the wording of an individual draft.

This is why meaningful validation cannot be reduced to a signature. The relevant question is what happened before approval. Did the reviewer inspect decisive evidence, consider contrary submissions, and independently assess the proposed interpretation?

Those are proposed evaluation questions, not findings about any particular court’s behavior. The reviewed materials do not establish that institutions have adopted the November framework. They show why its focus on inspectable justification has an existing institutional context.

The practical significance lies in making responsibilities more concrete. Developers can identify what reviewers need to see. Institutions can identify what an approval process should record. Researchers can then investigate whether those arrangements improve actual review.

Watch for Implementation, Independent Testing, and Court Adoption

The next meaningful developments would turn the conceptual framework into something that outsiders can examine and evaluate.

Over the coming months, the first signal to watch is a public implementation or detailed technical demonstration. It should expose how the proposed conditions operate on a worked dispute. A readable answer alone would provide limited evidence about the underlying specification.

The useful artifact would connect factual premises, legal authorities, objections, and conclusions. It should also identify unresolved issues and permit corrections. Such a release would strengthen the case that the framework can guide practical development.

A demonstration would still have limited reach. Selected examples can show feasibility without establishing reliability across unfamiliar matters. Readers should look for an explicit statement of coverage and avoid treating a narrow prototype as a general judicial reasoning engine.

The second signal is independent evaluation using difficult cases and deliberate errors. Evaluators should test missing evidence, conflicting rules, and decisive objections that a draft fails to address. These tests would examine whether the application exposes weaknesses instead of merely producing complete-looking documents.

Human performance should form part of that evaluation. Researchers need to examine whether reviewers notice more errors and understand the displayed reasoning. Review time also matters because an interface that overwhelms users can undermine its intended purpose.

A favorable result would strengthen the argument for inspectable representations in the tested setting. A disappointing result would identify limits in the implementation or its interface. Neither outcome should automatically settle every claim about judicial AI.

The third signal is documented institutional adoption with a defined scope. A court pilot, procurement requirement, or published review protocol would provide evidence of practical uptake. The documentation should identify what assistance is authorized and how responsibility operates.

Readers should distinguish those developments from general statements supporting transparency. A broad commitment does not establish use of this framework. An identifiable requirement, implementation, or evaluation would provide a firmer basis for assessing adoption.

These signals also clarify what remains unknown today. The publication does not establish accuracy gains, reduced review costs, or improved judicial outcomes. Those claims require evidence from systems and users working under specified conditions.

For developers, the immediate value is a more precise set of design questions. For institutional buyers, it is a way to interrogate what a product makes reviewable. For affected parties, the relevant issue remains whether the reasons behind a decision can be understood and challenged.

Read the original research with those distinctions in mind before evaluating the next judicial AI demonstration. Ask what evidence supports the displayed premises and which objections the system leaves unresolved. Check whether a reviewer can change the proposed interpretation without losing the supporting record. Look for independent results that distinguish drafting convenience from improved scrutiny. The Computable Due Process proposal supplies a vocabulary for that assessment, while implementation and testing remain ahead. When a vendor presents an AI-generated ruling, ask to inspect the reasoning connecting its evidence to its conclusion. Can a reviewer locate the precise step that deserves disagreement, and can the system preserve that challenge through revision?

Give every agent the context to do better work

Connect your agents to the knowledge, decisions, and history already organized in remio.

remio currently supports Windows 10+ (x64) and Macs with Apple silicon.

Your AI Partner at Work
Get more done with remio

Plan. Create. Deliver.
All in one place.

bottom of page