Computable due process: the guarantee of a reasoned decision as a specification of admissible legal reasoning in assisted judicial action

An artificial intelligence system can draft a text with every appearance of a reasoned judicial decision without containing the justification that ought to sustain it. The risk is a current one, since such systems already draft the document that the judge validates in assisted judicial action, and the law provides no precise criterion for telling effective reasoning apart from its appearance. This article supplies that criterion by reconstructing the constitutional duty to give reasons as a set of formal correctness conditions expressible in models of structured argumentation and defeasible deontic logic, which it terms computable due process . The method combines legal-dogmatic analysis with formal reconstruction, starting from Peruvian constitutional case law, which offers the most articulated typology of defective reasoning in the continental tradition, and whose six flaws correspond to five conditions on the argument: that it exist, that its inference be valid, that its premises be justified, that it meet the applicable standard of proof, and that it answer every claim. The threshold varies with the decision: a procedural order and a custodial conviction do not demand the same. That structure is not peculiar to Peru, since Brazil enacts six equivalent grounds by statute, Spain, the only legal order that has regulated assisted judicial action, grades the duty by type of decision, and the European and Inter-American courts require the same. Applied to a decided case, the model shows that Peruvian doctrine on circumstantial evidence already describes the components of a formal argument. Language models do not satisfy these conditions, since specialized legal tools with source retrieval produce false information between 17% and 33% of the time. They are especially exposed where a special norm clashes with a general one, since they may follow the one their training corpus repeats more often rather than the one the case requires. A design consequence follows: the burden of justification rests on an inspectable symbolic layer, the assessment on the judge. In conclusion, the criterion distinguishes artificial intelligence that sustains the judicial function from one that merely imitates it, and offers a verifiable pattern for the constitutional review of assisted decisions.