Claim-to-folio reproducibility audit of Király & Tokai (2018) on the Rohonc Codex: claim list, codebook, folio concordance, results and code

Version 1.2.0, 25 September 2026. This version replaces the faulty version 1.1.0 for readers. It contains the English research manuscript (version 0.8), carries the 23 September 2026 reporting corrections and updates the English documentation. Base data and stored numerical results are unchanged from version 1.0.0. Independent human review has not taken place; its absence is disclosed below. How much of a proposed reading of the Rohonc Codex can be checked from the open facsimile? This deposit accompanies a claim-to-folio audit of thirty location-specific claims in Király and Tokai (2018), Cracking the code of the Rohonc Codex. It provides the research manuscript, claim list, coding rules, folio concordance, results and original research scripts. The audit assesses the checkability of described features in the low-resolution monochrome facsimile of MTAK MS K 114; it does not assess whether the proposed decipherment is correct. AI-coded results. For the frozen thirty-claim list, the final AI-coded records contain 15 exactly reproducible claims, 14 partly reproducible claims, one unresolved locator and no claims classified as dependent on an unpublished sign decision. The exactly reproducible proportion was 0.50 (95% Wilson interval: 0.33–0.67). These are descriptive results for the selected sample. Start reading: the research manuscript Lux_2026_RohoncClaimAudit_manuscript_v0.8.pdf; the deposit guide README.md; the English codebook CODEBOOK_v1.0_EN_translation.md; the claim selection guide README_claims_v1.md; the folio reference guide README_folio_inventory.md. All are in the file list of this record. Research record. The plan and inputs were frozen before coding. The public OSF registration followed coding on 22 September 2026. The original Hungarian codebook and plan remain unchanged as historical records in version 1.0.0; the English guides provide the reader-facing documentation. The original scripts retain project-specific dependencies and have not been validated as a standalone executable package. Access and reuse. The complete archive Rohonc_v1.2.0_complete_deposit.zip preserves the metadata folder, the change record and file manifests. Neither the source facsimile nor the audited article PDF is redistributed; no manuscript image, crop or transcription is included. Text and data: CC BY 4.0. Python scripts: MIT. AI use. The project log records the following tool roles. Claude, through Claude Code (model identifier: claude-opus-5), performed claim selection, prepared the codebook, carried out both coding passes, and prepared the manuscript and release materials. Perplexity and ChatGPT performed external audits; their model identifiers are not recorded. Codex performed the historical Zenodo and OSF operations; its model identifier is not recorded. Veronika Lux initiated the project, made the source-access and publication-venue decisions, initiated external research runs and gave approvals. She did not code the claims and did not check the coding claim by claim. These recorded roles do not establish independent human validation. No AI system is listed as an author. OpenAI Codex assisted the English reader documentation. Version 1.2.0 (manuscript version 0.8, the deposit guide and the release documentation) was prepared by Claude through Claude Code (model identifier: claude-opus-5-5) on 25 September 2026; it changes presentation and documentation only, and the upload to Zenodo is done by Veronika Lux by hand. Review and validation status. This release has not undergone independent human peer review, and the claim codings have not been independently re-coded by a human. All codings were produced in AI-assisted coding runs. The six recorded repeat cases had identical verdicts and reason codes; this establishes agreement for those six stored outputs only and does not validate the coding decisions or establish consistency beyond those cases. The archived aggregation scripts reproduce the reported counts byte-identically from the stored coding outputs. This verifies the aggregation step only: it is not a reproduction from the source images, and it does not validate the codings. The reported figures describe 30 selected, location-bound claims from Király and Tokai (2018), assessed against the openly available low-resolution monochrome facsimile. They should not be generalized beyond this sample, and they are not evidence for or against any decipherment of the Rohonc Codex. No stored results changed relative to version 1.0.0.