Abstract
Fluent manuscript readings generated by artificial intelligence (AI) can obscure the distinction between visible evidence, supplied context, and reconstruction. This article proposes a publication provenance framework for AI-assisted research on Arabic-script Ibn Sina manuscripts. It specifies linked records for source identity, images, textual layers, generation and evaluation, uncertainty, specialist review, and revisions. The archive is treated as evidence about recorded project objects and events, not as self-validating evidence of the photographed manuscript text. We distinguish archival existence claims, explanatory claims about record mechanics, and comparative performance claims; the last are not supported by these materials. Twenty-four archived outputs associated with four witnesses illustrate the recording decisions without estimating or ranking system performance. The recovered scoring reports supply 432 criterion values, checked against 72 reported layer totals. These records clarify why an output with no substantive semantic coverage can receive a nonzero total: caution and editorial transparency contribute points independently of content recovery. Other cases show partial semantic agreement alongside meaning-changing reconstruction. Nominal Blind–Context pairs provide secondary provenance observations only. A crosswalk identifies existing support in TEI, IIIF, and PREMIS and distinguishes the proposed AI-specific recording policy. A worked example records known information, missing metadata, and unresolved textual status without implying verification. The contribution is a documentation and editorial workflow proposal whose interoperability, usability, and effects on scholarly decisions require independent evaluation.
1. Introduction
Historical manuscripts increasingly reach readers through digital images, transcriptions, translations, and searchable collections. These forms of access depend on editorial decisions that are not always visible in the final resource. Work on premodern Islamicate textual heritage has connected digitization with corpus construction and scholarly reuse (Kiessling et al., 2017; Miller et al., 2018). Research on classical Chinese medical texts likewise treats annotation, indexing, and dissemination as connected tasks (Liu et al., 2026). General-purpose multimodal AI adds another layer to this process: generated wording may combine what is visible in an image with contextual knowledge and plausible completion. A digital resource needs to preserve the origin and review status of each layer.
Ibn Sina’s works provide a demanding setting for this problem. Their transmission across medicine, philosophy, and logic involves technical vocabulary, textual variation, and translation between scholarly traditions. The long reception and monumental scale of the Canon of Medicine—which necessitated a vital medieval tradition of abridgements (such as al-Īlāqī’s al-Fusūl al-Ilāqiyya, Ibn al-Nafīs’s Mu‘jaz al-Qānūn, and al-Chaghmaynī’s Qānūncha) to make its vast contents accessible for daily medical practice and education—illustrate the importance of identifying the particular text and translation under discussion (Ibragimov, 2025; Veit, 2023).
1.1. Recognition and Publication Are Related but Distinct Tasks
Connected letterforms, variable dotting, damaged surfaces, marginal annotations, and mixed prose and verse complicate Arabic-script recognition. Historical Arabic scientific manuscripts have been addressed in dedicated recognition evaluations, including RASM2018 (Clausner et al., 2018). Recent work also includes KITAB-Bench for Arabic optical character recognition (OCR) and document understanding, InkSight for manuscript analysis and retrieval, and an evaluation of Gemini on naskh, taʿliq, and ruqʿah scripts (Adıgüzel & Özbek, 2026; Heakl et al., 2025; Sakhovskiy et al., 2026). These studies address recognition or system evaluation using their own datasets and procedures. Their findings cannot supply missing validation for a different archive.
The present article concerns the subsequent editorial question: what must remain attached to a generated reading before it is quoted, indexed, translated, or incorporated into a scholarly resource? Recognition accuracy alone does not document whether a phrase came from the image, an incipit supplied in a prompt, a translated reference, or later editorial correction. Existing standards already provide substantial support for manuscript description, image annotation, textual alternatives, responsibility, and preservation events. The proposed framework selects and connects those mechanisms for this particular workflow.
1.2. Collection and Scholarly Context
The Al-Biruni Institute of Oriental Studies in Tashkent preserves the witnesses discussed here. Its manuscript collection was inscribed in UNESCO’s Memory of the World Register in 1997 (UNESCO, n.d.). Ibn Sina materials are situated within a longstanding cataloguing tradition, including Sobranie vostochnykh rukopisei, Mirzayev’s bibliography, and Vakhabova’s dedicated catalogue (Mirzayev, 1955; Semenov et al., 1952–1975; Vakhabova, 1982). Shelfmarks and codicological descriptions establish the identity and context of a witness; they do not establish that an AI-generated passage accurately represents it.
The four selected witnesses concern logic, pharmacology, cardiac medicine, and didactic medical verse. Their digital surrogates range from small crops and screenshots to higher-resolution images. This heterogeneity makes them useful examples of the documentation problems encountered in research collections, without making them representative of the collection or of Arabic-script manuscripts generally.
1.3. Purpose, Epistemic Scope, and Contribution
The principal contribution is the framework in Section 3. The article asks three questions: which links between a manuscript image and a generated textual claim should a publication record retain; how can established standards support those links; and how should missing information and unresolved readings affect editorial use? Twenty-four previously archived outputs provide bounded illustrations of these questions in Section 4. All observations about generated content refer to those records.
This reframing changes the epistemic role of the archive. The archive documents what project files, outputs, scores, and metadata records exist and how they are linked; it does not by itself establish that an AI reading is textually correct, that a prompt condition caused an effect, or that one generator is superior. Its function in this article is evidentiary and documentary: to expose provenance, missingness, and the basis on which an editorial claim could later be reviewed. The contribution consists of a minimum record, a standards crosswalk, editorial decision rules, and a worked example that can subsequently be implemented and evaluated.
Three statement types are therefore kept distinct throughout. An existence statement reports an inspectable archival fact (for example, that a file, criterion value, metadata label, or checksum is present). An explanatory statement interprets relationships within those records (for example, that the rubric structure explains how a nonzero total can coexist with zero substantive semantic coverage); such explanations concern the documented record mechanics and are not claims about a model’s internal process. A comparative statement asserts relative performance, improvement, repeatability, or a causal effect across generators or prompt conditions. The present archive can support the first type and bounded instances of the second, but its design does not support the third.
2. Materials and Methods
2.1. Archival Materials and Nominal Prompt Conditions
The archive comprises four witness cases, three reported generator labels, and two nominal prompt-condition labels, with one output for each generator–condition–case combination: 24 outputs in total. Project metadata identify the generators as GPT-5.6-Sol at Max reasoning effort through ChatGPT, Gemini 3.1 Pro, and Claude Opus 4.8. These identifiers are retained as reported archival labels; complete platform exports are unavailable for independent verification of the selected builds and settings. Table 1 records the witnesses and digital material.
Table 1.
Manuscript witness cases and digital surrogates reported in the archive.
Terminology is fixed as follows. A witness is the catalogued manuscript object; a case is the project-level identifier (IS-01, SK-01, HL-01, or UR-01) linking a selected digital surrogate to its reference packet and archived outputs. Generator denotes the AI system that produced an output, whereas evaluator denotes the agent or system that assigned an assessment. Blind and Context are capitalized only as nominal archival condition labels: nominal Blind means that the intended prompt packet contained the image without a work or subject description, and nominal Context means that the intended packet also contained a work or subject description. These labels do not imply verified blinding or equivalent context. Stage 3 is the project-internal label for the post-generation assessment phase in which outputs were compared with the translated reference layers using the archived criterion tables; it is not a community standard or a new experiment introduced in this revision.
The witness descriptions derive from the project’s unpublished working inventory, Qisqacha tavsiflar [Brief descriptions of Ibn Sina manuscript witnesses], pp. 1–5, and the Institute’s catalogue resources (Al-Biruni Institute of Oriental Studies, Uzbekistan Academy of Sciences, n.d.). The working inventory is part of the project documentation described in the Data Availability Statement. Published Uzbek editions and studies provide editorial comparanda for the medical works (Hikmatullayev, 1966; Ibn Sina, 1987; Shoislomov, 1972). The folio ranges identify the catalogued works, not a verified line-level mapping of the evaluated images. The manuscript dimensions are codicological measurements; the pixel dimensions describe the submitted digital surrogates.
The generation guide called for a separate conversation for each case, generator, and nominal condition. The nominal Blind condition supplied the image; the nominal Context condition added a description of the work and subject. The protocol excluded reference translations, earlier outputs, external editions, dictionaries, and web searches. The reports describe heterogeneous images without common resampling or line segmentation. These are records of the intended procedure, rather than proof that every interface and session complied with it. Some nominal Context responses disclose prior nominal Blind exposure, and the Claude records describe catalogue information or Arabic incipits in the supplied context. The labels therefore remain archival descriptors throughout this article.
2.2. Reference Layers and Scope of Textual Assessment
The principal semantic references were human-prepared editorial translations in Uzbek Cyrillic. Academic-English working renderings, prepared with ChatGPT assistance, supplied an intermediate language for the reports. They were derived from the Uzbek references, not independently translated from the photographed Arabic. A diplomatic transcription records a witness’s written forms and uncertainties; an editorial translation already contains interpretation. The two should not be treated as interchangeable forms of ground truth.
The Stage 3 assessment reports divided the translated references into 12 semantic units for IS-01, 8 for SK-01, 12 for HL-01, and 12 for UR-01. The corresponding terminology lists contain 30, 13, 23, and 20 entries. These units support the archived meaning-level summaries. They do not provide Arabic token alignment. We use “reference conflict” for a discrepancy documented against these translated layers and reserve a claim of verified witness wording for a reading that has actually been checked against the manuscript.
2.3. Recovery, Auditability, and Validity of the Scoring Reports
For this revision, the four original case-specific Stage 3 assessment reports and the integrated report on Claude outputs were recovered. The case-specific reports identify System A as ChatGPT and System B as Gemini; the integrated report identifies System C as Claude. These archival labels identify the generators being assessed, not the evaluator. The three criterion tables in each report provide six values for normalized source reading, six for literal translation, and six for scholarly translation, each on a 0–5 scale. All 432 criterion values were transcribed from these tables. Their sums reproduce all 72 layer totals in the preceding manuscript. Supplementary File S1 provides the complete values, source filenames, table locations, and file fingerprints.
Two separate questions follow from this recovery. Auditability concerns whether a reader can trace a reported total back to a source file, criterion vector, textual layer, and arithmetic. Recovering all 432 criterion values, checking them against the 72 layer totals, and reporting filenames, table locations, and file fingerprints improves that traceability. Validity concerns whether the archived scores are defensible measures of manuscript-reading quality and can support inferential claims. That question depends on evidence not created by recovery, including independently verified reference text, known evaluator provenance, appropriate blinding, and validation of the scoring procedure for the intended construct. Accordingly, the recovery improves auditability but does not improve the evidential validity of the scores. The rubric also combines content recovery with transparency, uncertainty preservation, and control of unsupported additions; an abstaining response can therefore receive points while providing no substantive reading. The criteria remain archived judgments whose interpretation must be checked against coverage and documented errors. We do not retrospectively alter their weights or derive a publication threshold from their totals.
ChatGPT’s assistance with the English reference layer and the organization of Stage 3 assessment material is documented. The recovered scoring reports do not reliably identify the system or systems and exact versions that assigned the numerical values. We therefore cannot attribute every score to a named evaluator. They are retained as AI-assisted assessment records, not as independent specialist ratings. The case-specific reports also acknowledge that generator identities were visible in filenames, so the A/B/C labels did not establish effective blinding. ChatGPT’s involvement in reference preparation, while its outputs were also being assessed, creates a possible source of correlated judgment; direct self-evaluation cannot be confirmed or excluded. Research on LLM judges documents self-enhancement and other biases in different tasks (Zheng et al., 2023), but the presence or magnitude of such bias in this archive has not been measured.
The archived semantic-alignment percentage is 100 × (correct units + 0.5 × partially correct units)/reference units. Weighted terminology uses the same full- and half-credit rule with the number of terms as denominator. Supplementary File S1 reports the available component counts as well as the percentages. “Semantic coverage” below refers to this weighted concept-level proxy, not to the percentage of Arabic words correctly read. Model-expressed confidence is reproduced as a reported value; it is not an independently calibrated probability.
2.4. Development of the Framework
We organized the documented problems into record requirements: identify the witness and image, distinguish supplied information from recovered text, separate generated and edited layers, preserve the basis of evaluation, and record editorial responsibility. We then mapped these requirements to the official documentation of the Text Encoding Initiative (TEI P5), the International Image Interoperability Framework (IIIF Presentation API 3.0), and Preservation Metadata: Implementation Strategies (PREMIS 3.0). The crosswalk is a conceptual mapping of available mechanisms, not a schema-conformance test.
The worked example applies this recording policy to an existing HL-01 assessment. It distinguishes information present in the reports from information not documented there. No missing timestamp, evaluator identity, source reading, or approval is supplied retrospectively. Historical scores remain in Supplementary File S1; the main text uses selected records only to explain editorial decisions. The inferential and implementation limits are consolidated in Section 5.2.
3. Proposed Publication Provenance Framework
3.1. The Publication Record and Its Textual Layers
The proposed publication unit is a textual claim together with the evidence and interventions needed to interpret it. The record should link the manuscript witness, an identifiable image region, the exact generation inputs, the unchanged output, subsequent textual layers, and the relevant review decision. A case-level identifier alone is insufficient when different portions of one output have different evidential status.
Four textual layers should remain distinguishable. A diplomatic layer records the visible witness, including gaps and uncertain readings. A normalized layer records editorial regularization and links each change to the diplomatic form. A translation layer identifies the text translated, the translator or system, and whether the rendering is literal or interpretive. A commentary layer contains identification, historical explanation, and conjecture. A generated passage can propose material for any of these layers, but a fluent response does not establish that a diplomatic layer exists.
The record must also distinguish the availability of a datum from confidence in a reading. “Not recorded,” “not recoverable from the present archive,” “not applicable,” and “restricted” describe different metadata states. None should be encoded as a numerical zero or silently replaced with an estimate. Textual uncertainty, by contrast, should identify the affected passage and the person or system expressing it. These distinctions allow a resource to disclose its limits without conflating missing documentation with a measured failure.
3.2. Minimum Linked Record
Table 2 specifies a practical minimum for this workflow. The fields may be stored in linked files or repository records; they need not reside in one document. Identifiers should connect each derivative to its actual source. A file checksum can help detect replacement or corruption, but cannot authenticate the historical circumstances of generation or establish textual correctness.
Table 2.
Minimum linked record for AI-assisted manuscript publication.
Generation and evaluation should be separate events with separate agent records, even when the same system performs both. The evaluator record should identify the reference layer, rubric version, criterion values, input order, visibility of generator labels, and any human adjudication. This makes overlapping roles inspectable. It also prevents a file named after an evaluated system from being mistaken for evidence that the same system performed the evaluation.
3.3. Relationship to Established Standards
TEI already supports manuscript description, transcription and editorial alternatives, uncertainty, responsibility, software information, and revision history. IIIF supplies persistent resource identifiers, image presentation, and annotation targets on a Canvas. PREMIS supplies preservation objects, events, agents, and rights information. These are substantial existing capabilities, not innovations introduced by the present article. Table 3 separates them from the local policy decisions proposed for AI-assisted manuscript work (IIIF Consortium, 2020; PREMIS Editorial Committee, 2015; TEI Consortium, 2026).
Table 3.
Conceptual standards crosswalk and the local recording policy.
The additional contribution is the coordinated recording policy: retain exact generation and evaluation inputs, identify overlapping roles, attach criterion scores to their reference layer, and make specialist review status visible at the level of the proposed reading. Existing standards can express much of this information, but adopting a standard alone does not decide which records a project must preserve or when a generated passage may enter a scholarly textual layer. The implementation problem is therefore one of profiles, controlled vocabularies, cross-record identifiers, and repository workflow rather than simply selecting a standard.
The specific customization points correspond to the rows of Table 3. For TEI, the textual-intervention, uncertainty, and human-review records need a project profile (for example, an ODD-based customization) that constrains passage-level statuses such as generated proposal, specialist-verified, corrected, and unresolved, and links each status to responsible agents and revision events. For IIIF, the image-and-region record must identify the exact derivative or crop actually submitted and link it to external generation and evaluation provenance; a Canvas or region target alone does not record prompt, model, or assessment history. For PREMIS, the software/events and rights records require local event and role vocabularies for AI generation, AI-assisted evaluation, specialist adjudication, upload, retention, and redistribution decisions. PREMIS event provenance does not replace passage-level textual semantics, which remain linked to TEI and IIIF records.
Cross-standard implementation also requires stable shared identifiers and version rules so that one witness region, output, score row, and review decision remain traceable across systems. A practical difficulty is that repositories and commercial AI interfaces may not preserve the exact uploaded derivative, prompt packet, session state, model build, or retention settings; when such data are absent, the profile should encode explicit missingness rather than infer them. These are implementation customizations and constraints, not a claim that the core standards are deficient. Formal schema extension is not presumed unless a local system cannot express the required status or role distinctions, and no interoperable TEI, IIIF, or PREMIS implementation is claimed here.
3.4. Editorial Use and Release Decisions
Under the proposed workflow, raw outputs remain preserved as generated objects. Candidate readings receive a separate status from specialist-checked readings, and corrections create new versions rather than overwriting the original response. Review should identify the passage inspected, the reference image or other witness used, the specialist’s role, and the disposition of the claim. Useful dispositions include unresolved, accepted with stated uncertainty, corrected, and rejected. These are proposed local labels, not new standard vocabularies.
A passage may enter a verified transcription or translation layer only after a documented review appropriate to that layer. An accepted gap or uncertain reading can remain part of a scholarly edition; approval does not require false certainty or a complete text. Context-derived work identification may enter descriptive metadata with its own source, but should not be presented as image-recovered wording. If a response contains only a probable formula and omits the substantive passage, its annotation can be retained while the corresponding transcription field remains empty or explicitly unresolved.
Discovery systems should preserve these distinctions. Generated suggestions may be searchable as clearly labelled suggestions, while verified text is indexed under its reviewed status. A record with unknown image-use permissions also needs a separate rights decision before dissemination. Textual approval, permission to upload an image to a commercial platform, and permission to redistribute it are different decisions. The proposed workflow records each without inferring one from another.
3.5. Worked Example from HL-01
Table 4 shows how the framework records the Claude nominal Blind assessment for HL-01. It is a retrospective documentation example based on the recovered reports, not a new assessment of the Arabic image. The local identifier introduced here connects the criterion row in Supplementary File S1 to this discussion; it is not an original platform run identifier.
Table 4.
Retrospective documentation example for HL-01 Claude, nominal Blind.
The example shows what the framework changes in practice: it makes the distinction between a numerical rubric total, the content actually reported as recovered, and a publication decision explicit. The available record supports retaining an illustrative annotation about abstention. It does not support entering a physiological passage as verified text. Recovering the missing generation or evaluation log would improve the record; independently checking a reading would require an additional scholarly event.
4. Illustrative Archival Cases
4.1. Why a Rubric Total Does Not Measure Usable Text
The HL-01 nominal Blind record has a normalized-source-reading total of 15/30, with criterion values 1, 0, 0, 4, 5, and 5. Fourteen points come from editorial transparency, preservation of ambiguity, and absence of unsupported reconstruction; one point comes from lexical normalization. The same record reports 0% semantic coverage, 0% weighted terminology, and only a probable basmala. It therefore illustrates a mixed-construct total: most points concern how the response handles uncertainty, while none establish recovery of the physiological content. The recovered criterion values resolve the ambiguity left by the total alone. This is an explanatory claim about the composition of an archived score, not a comparative claim about generator quality.
The IS-01 abstention records provide a related example: normalized-source-reading totals of 13/30 coexist with zero semantic coverage. Reporting the criterion vector beside the total makes this interpretable, but does not validate the rubric as an interval scale or justify comparison between systems. The framework consequently requires both the assessment record and the actual textual contribution. No “usable text” threshold is inferred from a 30-point total.
4.2. Partial Agreement and Documented Reference Conflicts
Other records illustrate the opposite editorial problem. The Gemini SK-01 assessments report 62.50% semantic alignment and 80.77% weighted terminology, yet document a reversal in the relation between vinegar and honey and confusion between attenuation and cleansing. These historical pharmacological propositions are objects of textual analysis, not medical recommendations. A high rate of term recognition cannot establish that the relationships among those terms have been preserved.
In UR-01, the Gemini and Claude assessments report 50.00% weighted semantic alignment. The reports nevertheless identify displacement of the reference’s sequence concerning nature, virtue, craft, language, poet, and physician by other material. The ChatGPT record has lower reported coverage and visible gaps, but also contains an unsupported proposed blessing. These examples support segment-level review; they do not establish a general preference for a system or for either completion or abstention. Table 5 connects the case records to specific documentation decisions.
Table 5.
Selected case records and the requirements they illustrate.
We retain these discrepancies as findings of the archived assessment reports. Without line-linked Arabic reference text and independent specialist adjudication, the present documentation exercise cannot determine the exact token-level error distribution or resolve every possible difference between a witness and an editorial translation. The practical consequence is to preserve the disputed passage and the basis of the report, rather than convert a reported conflict into a newly verified Arabic correction.
4.3. Prompt History and Expressed Confidence
The nominal Blind–Context pairs supply secondary descriptive observations about documentation. For SK-01, title-level additions in the ChatGPT and Claude nominal Context records coexist with zero substantive semantic coverage. These additions cannot demonstrate an improvement in image-based reading when related information may have been supplied in the prompt. Gemini’s paired scores and principal findings are identical across the four cases, but the independence of the paired generations and evaluations cannot be verified. Their agreement is not evidence of repeatability or of an absence of a context effect. No comparative performance claim is drawn from these pairs.
Confidence requires a similarly bounded interpretation. The Gemini UR-01 record reports 94% confidence, while the archived assessment documents a major reference conflict. The Claude UR-01 record reports 70%; the HL-01 Claude nominal Blind record reports approximately 1%. The object and elicitation of confidence were not standardized, so these numbers should not be compared as calibrated probabilities. The framework retains the original confidence statement, its object, and the associated disputed reading. It does not turn expressed confidence into editorial approval.
5. Discussion
5.1. Contribution to Manuscript Publishing
The framework addresses a recurring transition in digital scholarship: provisional machine output can acquire the appearance of an edited source once it is copied into a catalogue, index, translation, or edition. Describing a passage as plausible, fluent, or reference-aligned does not record its origin or the interventions through which it reached publication. The proposed links make those interventions inspectable.
This contribution is narrower than a new metadata standard and different from a validated evaluation instrument. It brings existing standards into a concrete recording and review policy for AI-assisted manuscript research. Its main features are the separation of textual layers, explicit generation and evaluation roles, linkage of scores to references and criteria, and passage-level review status. The recovered score tables make one particular archive more transparent, while the worked example demonstrates how the policy records both available information and unresolved gaps.
The same policy can accompany outputs from dedicated recognition software or general-purpose assistants. The present archive offers no evidence of superiority over a handwritten text recognition (HTR) or OCR system. Recognition benchmarks and a provenance workflow answer different questions: a benchmark evaluates a specified task under defined conditions, whereas the workflow documents the basis and permitted interpretation of a particular textual object. Their combination would allow future projects to retain both measurement and editorial accountability.
5.2. Limitations
The archive contains four purposively selected witness cases and one generation per generator–condition–case combination. It cannot estimate stochastic variation, population performance, or general model characteristics. Image quality, script, genre, and content vary together. Incomplete session logs, reported prior exposure, non-equivalent context packets, and unverified independence of paired Gemini records prevent causal interpretation of the nominal conditions.
The reference layers are Uzbek editorial translations and AI-assisted English renderings, without line-linked Arabic diplomatic ground truth. They support bounded meaning-level discussion but cannot establish character error rate, word error rate, token-level hallucination rates, or a calibrated probability of correct source reading. The project-internal Stage 3 assessment reports contain AI-assisted judgments without independent blinded specialist validation, and their numerical evaluator identities remain incompletely documented. Recovering all criterion values and checking their arithmetic improves auditability only; it does not remedy these validity limitations.
The proposed framework has not been implemented as a standards-conformant system or tested with independent editors or repositories. Its field selection and review rules remain proposals. The archive also does not establish the specific permissions for commercial-platform image uploads or the retention and training settings used at the time. Access to a repository image should not be interpreted as documenting those permissions. Supplementary File S1 distinguishes recovered assessment records from information still unavailable; it is not a complete executable replication package.
5.3. Evaluation Needed Before Wider Adoption
A subsequent implementation study should test whether independent users can locate the source image, distinguish generated from verified wording, identify the responsible agents, and follow a correction across versions. Interoperability should be checked against explicit TEI, IIIF, and PREMIS implementations rather than inferred from a conceptual crosswalk. These tests would evaluate the framework itself.
A separate empirical study of manuscript reading would require a broader and justified sample, line-linked Arabic references, standardized inputs, preserved platform and session logs, and repeated generations chosen according to the precision sought. Independent specialists should apply a prespecified rubric and adjudication procedure, with agreement and uncertainty reported. Generator and evaluator roles should be recorded separately, and any overlap examined. These requirements are a future evaluation design, not work completed in the present revision.
6. Conclusions
This article proposes a publication provenance framework for AI-assisted research on Arabic-script Ibn Sina manuscripts. It specifies what to retain about the source image, textual layers, generation and evaluation events, uncertainty, specialist review, and revision history. Its standards crosswalk shows that many required mechanisms already exist in TEI, IIIF, and PREMIS; the contribution is a coordinated policy for applying them to generated manuscript readings. Practical implementation nevertheless requires local profiles or controlled vocabularies, stable cross-standard identifiers, and explicit treatment of missing platform metadata rather than assuming that a conceptual mapping is already interoperable.
The 24 archived outputs provide illustrative cases only. The archive supports existence claims about recorded project objects and bounded explanatory claims about score composition and provenance. It does not support comparative performance claims, model rankings, or a causal effect of context. The recovered criterion scores clarify how nonzero totals can coexist with no substantive text, while other records show how partial semantic agreement can coexist with a meaning-changing reference conflict. These observations justify keeping coverage, uncertainty, and documented review distinct in the proposed record. Independent implementation and user evaluation remain necessary to determine whether the framework improves scholarly review and publication practice.
Supplementary Materials
The following supporting information can be downloaded at: https://www.mdpi.com/article/10.3390/publications14040080/s1, Supplementary File S1, Archival Scores and Publication Provenance, accompanies this article. It contains the material and source-report inventories, all 432 criterion values and 72 layer totals, semantic and terminology summaries, available component counts, source-report fingerprints, the worked record, and remaining documentation limits. Table S1. Material inventory and reference denominators. Table S2. Sources of the criterion values. Table S3. Criterion code key; all individual values range from 0 to 5. Table S4. Normalized source reading criterion values and reproduced totals. Table S5. Literal translation criterion values and reproduced totals. Table S6. Scholarly translation criterion values and reproduced totals. Table S7. Archived totals alongside semantic alignment, terminology, and confidence. Table S8. Available component counts behind the semantic and terminology summaries. Table S9. IS-01 translated-reference unit descriptions. Table S10. SK-01 translated-reference unit descriptions. Table S11. HL-01 translated-reference unit descriptions. Table S12. UR-01 translated-reference unit descriptions. Table S13. Recovered information and remaining documentation limits. Table S14. Worked provenance record for HL-01-C-B.
Author Contributions
Conceptualization, B.A. and F.N.I.; methodology, F.N.I. and A.A.u.I.; validation of manuscript interpretation, B.A. and F.N.I.; formal analysis, A.A.u.I.; investigation, F.N.I. and A.A.u.I.; resources, B.A. and F.N.I.; data curation, A.A.u.I.; writing—original draft preparation, A.A.u.I.; writing—review and editing, B.A., F.N.I. and A.A.u.I.; visualization, A.A.u.I.; supervision, B.A. and F.N.I.; project administration, B.A. Validation here refers to authorial review of the interpretation and does not denote independent blinded scoring of the archived outputs. All authors have read and agreed to the published version of the manuscript.
Funding
This research received no external funding.
Data Availability Statement
The criterion values, archived layer totals, semantic and terminology summaries, and documentation example used in this article are included in Supplementary File S1. The five source assessment reports are identified there by filename, table location, and SHA-256 fingerprint. The manuscript witnesses are held in the Main Fund of the Al-Biruni Institute of Oriental Studies under shelfmarks 2213, 2572/XXI, 2275/V, and 3008/VI. The working witness descriptions, generation guide, reference translations, raw outputs, and source reports are retained in the project archive and may be requested from the corresponding author for scholarly or editorial verification, subject to applicable repository and reproduction permissions. Full-resolution manuscript images are not redistributed in the supplement. The package does not contain complete platform session exports or independently verified Arabic diplomatic transcriptions.
Acknowledgments
This research was conducted within the work plan of the Department of History of Science and Digital Manuscripts of the Al-Biruni Institute of Oriental Studies, Uzbekistan Academy of Sciences. The authors thank the Institute for access to manuscript descriptions and digital research materials. Use of Generative AI: OpenAI ChatGPT assisted with academic-English renderings of author-supplied Uzbek editorial references, the organization of project-internal Stage 3 assessment material, language and structural editing, bibliographic cross-checking, extraction and consistency checking of archived score tables, and document preparation. ChatGPT, Gemini, and Claude were also the generators whose archived outputs form the illustrative cases. Stage 3 assessments are AI-assisted records, and the available documentation does not establish the numerical evaluator system and version for every score. No new model generations, specialist ratings, or Arabic diplomatic ground-truth transcriptions were produced for this revision. The authors are responsible for the submitted interpretation, citations, and disclosure of AI use.
Conflicts of Interest
The authors declare no conflicts of interest.
References
- Adıgüzel, A., & Özbek, M. S. (2026). Digital paleography in Arabic manuscripts: The performance levels of the Gemini model in recognizing Naskh, Ta‘liq, and Ruq‘ah scripts. Dinbilimleri Akademik Araştırma Dergisi, 26(2), 909–941. [Google Scholar] [CrossRef] [Scilit]
- Al-Biruni Institute of Oriental Studies, Uzbekistan Academy of Sciences. (n.d.). Elektron katalog bazasi [Electronic catalogue portal]. Available online: https://beruni.uz/en/9-explore/256-elektron-katalog-bazasi.html (accessed on 10 November 2026).
- Clausner, C., Antonacopoulos, A., McGregor, N., & Wilson-Nunn, D. (2018). ICFHR 2018 competition on recognition of historical Arabic scientific manuscripts—RASM2018. In 2018 16th international conference on frontiers in handwriting recognition (ICFHR) (pp. 471–476). IEEE. [Google Scholar] [CrossRef] [Scilit]
- Heakl, A., Sohail, M. A., Ranjan, M., Elbadry, R., Ahmad, G. S., El-Geish, M., Maher, O., Shen, Z., Khan, F. S., & Khan, S. (2025). KITAB-Bench: A comprehensive multi-domain benchmark for Arabic OCR and document understanding. In Findings of the association for computational linguistics: ACL 2025 (pp. 22006–22024). Association for Computational Linguistics. [Google Scholar] [CrossRef] [Scilit]
- Hikmatullayev, H. (1966). Abu Ali ibn sinoning “Yurak dorilari” nomli risolasi [Abu Ali Ibn Sina’s treatise “Cardiac drugs”]. Fan. Library catalogue record. Available online: https://book.iiau.uz/book?id=2494 (accessed on 10 November 2026).
- Ibn Sina, A. A. (1987). Tibbiy risolalar [Medical treatises] (U. I. Karimov, & H. Hikmatullayev, Eds.; Vol. 1, p. 145). Fan. National bibliography record, entry 1065. Available online: https://uzbekbibliography.library.illinois.edu/Content/PDFs/1987.pdf.ocr_ezhegodnikknigi.pdf#page=145 (accessed on 10 November 2026).
- Ibragimov, F. N. (2025). Dissemination of medical knowledge: The role of al-Qanun abridgements in the Muslim East. Research on History of Medicine, 14(Suppl. 1), S7–S10. [Google Scholar] [CrossRef]
- IIIF Consortium. (2020). IIIF presentation API 3.0. Available online: https://iiif.io/api/presentation/3.0/ (accessed on 10 November 2026).
- Kiessling, B., Miller, M. T., Romanov, M. G., & Savant, S. B. (2017). Important new developments in Arabographic optical character recognition. Al-ʿUṣūr al-Wusṭā, 25, 1–13. [Google Scholar] [CrossRef]
- Liu, S., Zhou, Y., Song, Y., & Chen, C. (2026). Methods, challenges, and future directions in annotation and indexing of classical Chinese medical texts: A narrative review. Publications, 14(2), 30. [Google Scholar] [CrossRef] [Scilit]
- Miller, M. T., Romanov, M. G., & Savant, S. B. (2018). Digitizing the textual heritage of the premodern Islamicate world: Principles and plans. International Journal of Middle East Studies, 50(1), 103–109. [Google Scholar] [CrossRef] [Scilit]
- Mirzayev, S. (1955). Ibn Sinoning Sharqshunoslik institutda mavjud asarlari: Bibliografiya [Works of Ibn Sina held at the Institute of Oriental Studies: A bibliography] (p. 382). National bibliography record, entry 3667. Academy of Sciences of the Uzbek SSR. Available online: https://uzbekbibliography.library.illinois.edu/Content/PDFs/1951-55.pdf.ocr_kniga_sovietskogoBW.pdf#page=382 (accessed on 10 November 2026).
- PREMIS Editorial Committee. (2015). PREMIS data dictionary for preservation metadata (Version 3.0). Library of Congress. Available online: https://www.loc.gov/standards/premis/v3/ (accessed on 10 November 2026).
- Sakhovskiy, A., Ulitin, I., Bojarskaja, E., Kokh, V., Murtazin, R., Novopoltsev, M., & Budennyy, S. (2026). InkSight: Towards AI-aided historical manuscript analysis. In Proceedings of the 19th conference of the european chapter of the association for computational linguistics (volume 3: System demonstrations) (pp. 271–281). Association for Computational Linguistics. [Google Scholar] [CrossRef] [Scilit]
- Semenov, A. A., Voronovskii, D. G., Urunbaev, A. U., & Epifanova, L. M. (Eds.). (1952–1975). Sobranie vostochnykh rukopisei Akademii Nauk Uzbekskoi SSR [Collection of Oriental manuscripts of the Academy of Sciences of the Uzbek SSR] (Vols. 1–10). Editors vary by volume. Fan. Series catalogue record. Academy of Sciences of the Uzbek SSR. Available online: https://ci.nii.ac.jp/ncid/BA06834785 (accessed on 10 November 2026).
- Shoislomov, S. (1972). Ibn Sinoning tib haqidagi she’riy asari “Urjuza” [Ibn Sina’s poetic work on medicine “Urjuza”] (U. I. Karimov, Ed.; p. 287). Fan. National bibliography record, entry 2811. Available online: https://uzbekbibliography.library.illinois.edu/Content/PDFs/1970-1972.pdf.ocr_kniga_sovietskogoBW.pdf#page=287 (accessed on 10 November 2026).
- TEI Consortium. (2026). TEI P5: Guidelines for electronic text encoding and interchange (Version 4.12.0). Available online: https://tei-c.org/Vault/P5/4.12.0/doc/tei-p5-doc/en/html/index.html (accessed on 10 November 2026).
- UNESCO. (n.d.). The collection of the Al-Biruni institute of oriental studies. Memory of the World. Available online: https://www.unesco.org/en/memory-world/collection-al-biruni-institute-oriental-studies (accessed on 10 November 2026).
- Vakhabova, B. (1982). Rukopisi proizvedenii Ibn Siny v sobranii instituta vostokovedeniia AN UzSSR [Manuscripts of Ibn Sina’s works in the collection of the Institute of Oriental Studies of the Academy of Sciences of the Uzbek SSR]. Fan. Library catalogue record. Available online: https://ci.nii.ac.jp/ncid/BA43267210 (accessed on 10 November 2026).
- Veit, R. (2023). Materia medica in a multilingual context: Avicenna’s Canon of medicine and its Latin translation of book II. Medieval Encounters, 29(2–3), 196–221. [Google Scholar] [CrossRef] [Scilit]
- Zheng, L., Chiang, W.-L., Sheng, Y., Zhuang, S., Wu, Z., Zhuang, Y., Lin, Z., Li, Z., Li, D., Xing, E. P., Zhang, H., Gonzalez, J. E., & Stoica, I. (2023). Judging LLM-as-a-judge with MT-bench and chatbot arena. Advances in Neural Information Processing Systems, 36, 46595–46623. [Google Scholar] [CrossRef] [Scilit]
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content. |
© 2026 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license.