Introduction

Investigative practice generates computational outputs of radically different kinds. A cryptographic hash establishes that a copied disk image is identical to its source; a one-to-many search returns ranked candidate faces above a similarity threshold; a model scores the probability that a person will abscond. Each enters the file as a document, and legislation treats none expressly. Doctrine has answered globally — assimilating such outputs to documentary or expert evidence, or excluding them as unverifiable — and both answers fail alike, by classifying according to technology rather than to the structure of the reasoning that supports the output.

For Uzbekistan the question is concrete: the Criminal Procedure Code now names digital evidence as a species alongside material and written evidence, Law No. LRU-1003 of 21.11.2024 extended forensic examination to information systems and data on Internet resources, and the «Digital Prosecution — 2030" Strategy provides for «E-tergov» with express contemplation of artificial intelligence. National doctrine confirms the gap rather than filling it: Imomnazarov treats the examination of digital data as an investigative action, neither reaching admissibility [1]. The purpose is to establish a criterion determining the procedural status of such results; the methods are formal-legal, comparative-legal and international-legal, with systemic analysis and modelling.

Two limitations are declared. The three-moment structure of control used below — before, during and after deployment — is not original: it is established in the literature on technological due process formulated by Citron and extended by Citron and Pasquale and by Crawford and Schultz, and it is the architecture of the Artificial Intelligence Act itself, combining conformity assessment, logging and rights of complaint and explanation [2]. What is contributed is the content assigned to each moment. The second limitation is empirical: this is a doctrinal study containing no original data on Uzbek practice, such data being unavailable.

I. Justificatory structures

The Artificial Intelligence Act defines such a system by reference to one that infers, from the input it receives, «how to generate outputs such as predictions, content, recommendations, or decisions» influencing physical or virtual environments [3]. This isolates the operative feature — something is produced upon which a person may act — but groups mechanisms whose claims to reliability rest on different grounds. Expert systems reason from rules stated in advance, so their justification is deductive and inspectable; machine learning derives its decision function from training data, so the warrant for an individual output is aggregate measured performance rather than a statement of why this case was decided so, and deep learning intensifies this into the opacity Pasquale termed the black box. The characterisation of big data by volume, velocity and variety originates in Laney's model of 2001, but the legally significant property is a fourth: analytical yield arises from combining data gathered for other purposes [4]. Digital forensics stands apart: the sequence named in the title of ISO/IEC 27037 — identification, collection, acquisition and preservation — with ISO/IEC 27041–27043 and accreditation under ISO/IEC 17025, is designed for repetition on the same object by another examiner.

The operative distinction is therefore not between traditional and algorithmic methods — forensic examination is itself tool-dependent — but between two justificatory structures. In the first, reliability is established by a procedure applied to this object and repeatable on it; in the second, by the statistical performance of a model over a population to which the object has been assigned. The first supports a claim about the case, the second a claim about a class from which a claim about the case is inferred. Classification must accordingly attach to the structure, not the technology, since the same technique may be embedded in either.

That algorithmic outputs raise a distinctively epistemic problem requires demonstration in the vocabulary of evidence theory. Damaška showed that rules of evidence are artefacts of an institutional arrangement of adjudicative authority, and identified the scientisation of proof as displacing epistemic authority from the trier of fact towards institutions whose competence the court cannot appraise; Ho Hock Lai developed the correlative duty, that the fact-finder asserts a proposition for which he is personally answerable and can do so only on grounds he can own and state; Allen and Pardo describe proof as a comparison of competing explanations rather than aggregation of probabilities over isolated items [5].

The two structures behave differently here. A reproducible case-specific examination furnishes material that enters an explanatory account and can be tested within it: the defence may contest the acquisition, commission an independent examination of the same object, or advance an alternative explanation of the same trace. A population-level attributive output enters no explanatory account; it deliversa conclusion that explains nothing and cannot be tested by re-examining an object, there being no object but a classification. Two objections make this precise: the reference-class problem establishes that a probabilistic statement about an individual varies with the class into which he is placed, no legal criterion determining that choice, so the modeller's selection performs work the law has not authorised; and the sensitivity objection explains why merely statistical evidence, however accurate in the aggregate, fails to track truth counterfactually, since the output would have been identical had the person been innocent, provided he shared the class features [6]. The argument thus condemns not computational methods generally but one inferential move — from class membership to individual attribution. It departs accordingly from Roth, who assimilates instrument outputs to conveyances requiring an analogue of confrontation, and from Quattrocolo, who doubts such results can satisfy adversarial testing at all; Gless's process-orientation is accepted but narrowed, the relevant process being not the software lifecycle in general but the inferential warrant of the particular output [7].

Three regimes follow. Forensic-instrumental outputs — hash verification, documented acquisition, extraction reports, computer-technical and comparative identification examinations conducted on a stated methodology — are evidence, governed by the ordinary rules on documents and expert conclusions supplemented by a validation triad: traceable provenance of the input data; a method validated with a disclosed error rate under stated operating conditions; and reproducibility by an independent examiner. The triad is adjacent to but not identical with the Daubert criteria, which address the method in the abstract; it adds the lineage of the specific data and the availability of independent repetition. Investigative-heuristic outputs — link analysis, anomaly detection, retrieval of similar cases and candidate lists from biometric searches — are orientation information: they may direct investigative activity but cannot ground a finding of fact, and their use must be recorded.

The distinction most often collapsed lies here. A one-to-many search is a database operation returning ranked candidates under a threshold; it is a lead. A comparative portrait or phonoscopic examination is a different procedure: an expert applies a stated methodology to two specified samples, declares error rates, and answers for the conclusion. Same technology, different justificatory structures, therefore different status — a distinction confirmed by wrongful arrests following treatment of a candidate return as an identification, and by error rates varying across demographic groups [8]. Predictive-attributive outputs — risk of absconding, recidivism scores, individualised crime prediction — are excluded as grounds for decisions concerning the person, by reason of the two objections rather than of inaccuracy, in correspondence with Article 5 of the Artificial Intelligence Act and with State v. Loomis [9]. The exclusion being categorical, the right of exclusion acquires a determinate object: an item in the first category may be excluded for failure of the triad, an output of the third is inadmissible as such.

II. International standards and comparative allocation

International instruments supply not a classification of outputs but a structure of control at three moments. At authorisation, Article 15 of the Convention on Cybercrime and Article 24 of the United Nations Convention against Cybercrime subject procedural powers over computer data to safeguards including independent supervision and limitation of scope; but the decisive development is judicial — the German Federal Constitutional Court held on 16 February 2023 that the algorithmic combination of police datasets is an interference distinct from the collection of the underlying data, striking down the enabling provisions for want of a proportionate threshold and of limits on the sources combined. That analysis, and not only collection, is independently regulated is the most consequential result in this field. At admission the standards are technical, the ISO/IEC series supplying provenance and reproducibility criteria and the NIST evaluations error rates disaggregated by group. At contestation, the Council of Europe Framework Convention on Artificial Intelligence, open to non-member States and thus available to Uzbekistan, requires effective remedies and the possibility of contesting decisions [10].

Judicial practice supplies content the instruments lack. In Glukhin v. Russia the European Court found the processing of biometric data by facial recognition an interference with private life and its use against a peaceful demonstrator incompatible with the values of a democratic society; one passage bears directly on the proposals below — the Court noted the applicant's difficulty in proving that the technology had been used, since domestic law provided no official record or notification of its use. Absence of a recording obligation is therefore not a technical omission but a defect disabling the remedy. In SyRI the District Court of The Hague declared the enabling legislation non-binding for failure to strike a fair balance under Article 8(2), the decisive consideration being that the system was insufficiently transparent and verifiable. The controversy between Wachter and her co-authors and Selbst and Powles over a right to explanation is resolved in practice in favour of contestability — hence the framing of the triad in terms of reproducibility rather than interpretability [11].

Comparison requires a stated tertium comparationis; the criterion adopted is the allocation of the burden of validation among executive, court and defence at each of the three moments, qualified twice: an allocation rule transfers only if the institution assigned the burden can discharge it, and the Watson–Legrand dispute is resolved intermediately, the rule transferring while the capacity to bear it must be separately built. Germany allocates authorisation to the legislature and constitutional court and regulates the electronic file in the Code itself, mandatory from 2026 though with a transitional derogation — the rule transferable, the derogation a warning that deadlines without infrastructure produce derogations rather than compliance. The United Kingdom allocates admission to the executive through statutory forensic accreditation but fails at contestation, where the volume of digital material has repeatedly defeated disclosure. France prohibits decisions assessing a person's conduct from resting on automated processing, and its time-limited authorisation with compulsory evaluation is a transferable technique. Estonia allocates part of the contestation burden to the individual by making the log of queries accessible to him — the most economical solution encountered, whose precondition exists in Uzbekistan in functional form. Singapore allocates almost everything to the executive; Korea is instructive for sequence; China yields no transferable rule [12]. Measured on the stated criterion, no order examined is satisfactory at all three moments: authorisation is best developed in Germany, admission in the United Kingdom, contestation in Estonia, and nowhere are the three combined — a legislature acting now composes a model rather than receives one.

IV. The legislation of the Republic of Uzbekistan

Four deficiencies follow from the texts. First, digital evidence has been introduced as a species without a criterion of admissibility: Chapter 25 of Criminal Procedure Code regulates attachment to the case, not the conditions under which reliability is established, so a heuristic search return and a forensic conclusion may enter the file with indistinguishable status. Second, the combination of data held in different state information systems for investigative purposes is not constituted as an act requiring authorisation — the deficiency identified in Germany in 2023. Third, admissibility is not conditioned on validation: the accreditation requirement of Resolution of the Cabinet of Ministers of the Republic of Uzbekistan No. 849 of 18.12.2024 operates as an administrative duty of the expert institution, not as a condition of the evidentiary use of its conclusions, and no norm requires disclosure of error rates. Fourth, there is no recording obligation and therefore no point of attachment for contestation — the defect identified in Glukhin as disabling proof of the interference itself.

The limits of this assessment must be stated: these are conclusions about the legal framework, not about practice. Whether algorithmic tools are in fact used in Uzbek pre-trial proceedings, and with what effect, cannot be established from open sources, the only relevant published judicial data concerning review-instance outcomes generally [13]. Closing the gap requires a definable programme: a sample of terminated cases coded for the type of digital evidence and the fate of each item; departmental statistics on computer-technical and phonoscopic examinations, including inconclusive results and laboratory accreditation; records of defence applications; and interviews with practitioners. The deficiency is thus not absence of policy but absence of procedural translation: transparency and human oversight are proclaimed in the strategic and ethical acts but acquire legal force only as rules of authorisation, admissibility and contestation in the Code.

V. Proposals

First, introduce into the Criminal Procedure Code the three categories with their differing consequences: forensic-instrumental results as evidence subject to validation; heuristic results as orientation information, sufficient to found further investigative action but not a finding of fact, and subject to mandatory recording; attributive predictive results as inadmissible grounds for decisions concerning the person. Second, establish the validation triad as a condition of admissibility in the first category, converting the accreditation requirement of Resolution No. 849 from an administrative duty into a condition of the evidentiary use of expert conclusions. Third, constitute the combination of data from two or more state information systems for investigative purposes as an independent procedural action requiring the sanction of a prosecutor or court, proportionate to the gravity of the offence and limited in time and in the registers involved. Fourth, create the point of attachment for contestation: an obligation to record every use of an algorithmic system, with rights of the defence to obtain the validation documentation, to seek an independent examination and to apply for exclusion. These proposals are mutually conditioning: categorisation without a validation test yields labels without consequences, a validation test without a recording obligation cannot be invoked, and a right of contestation without technically competent counsel is a right on paper.

Conclusion

The procedural status of a computational output depends on its justificatory structure rather than on the technology producing it, and three structures must be distinguished: case-specific reproducible validation, heuristic search, and population-level attribution. The epistemic objection applies to the third alone; extending it to the first would destroy the evidentiary value of digital forensics. The theoretical contribution is a criterion of classification derived from evidence theory rather than technology, showing the asserted incompatibility between algorithmic methods and criminal proof to be partial and specifiable; the practical contribution is its formulation as norms capable of incorporation into the Criminal Procedure Code and into the specification of «E-tergov».

References:

1. Imomnazarov A. Abstract of dissertation, 12.00.09. Tashkent, 2025;
2. Citron D. K., Pasquale F. // Washington Law Review. 2014. Vol. 89, No. 1. P. 1–33;
3. Regulation (EU) 2024/1689, art. 3(1), following OECD, OECD/LEGAL/0449 (2019, rev. 2024);
4. Pasquale F. The Black Box Society. Harvard University Press, 2015; Laney D. 3D Data Management. META Group Research Note, 6 February 2001;
5. Allen R. J., Pardo M. S. // International Journal of Evidence & Proof. 2019. Vol. 23, No. 1–2. P. 5–59;
6. Enoch D., Spectre L., Fisher T. // Philosophy & Public Affairs. 2012. Vol. 40, No. 3. P. 197–224;
7. Roth A. // Yale Law Journal. 2017. Vol. 126, No. 7. P. 1972–2053;
8. NIST, NISTIR 8280 (2019);
9. Regulation (EU) 2024/1689, art. 5(1)(d);
10. Convention on Cybercrime (ETS No. 185), art. 15;
11. Glukhin v. Russia, no. 11519/20, 4 July 2023, §§ 72, 73, 90;
12. Watson A. Legal Transplants. 2nd ed. 1993;
13. Resolution of the Cabinet of Ministers No. 849 of 18 December 2024.