errata-check: Deterministic auditing of errata against frozen published artifacts / 凍結された公開物に対する、決定的な正誤表の監査

Abstract

A published artifact with a DOI cannot be revised; only the errata can. Errata therefore drift away from the primary source over time: a quotation loses a word, occurrences are undercounted, an item declared unresolvable is quietly marked resolved, and the last-updated stamp goes stale. This tool turns that drift into a failing check in continuous integration. It verifies that every passage the errata claims is printed does appear verbatim in the source PDF; that passages the errata claims are absent really are absent; that the declared number of occurrences matches; that files claimed to be bundled are genuinely missing; that counts named in prose match what the corresponding command actually reports; that items declared unresolvable still say so; that the frozen artifact itself has not been swapped (SHA-256); and that the last-updated date is not older than any date inside the document. No inference is used in the judgement. No large language model, no similarity score. A string is present or it is not; a digest matches or it does not. The output therefore does not need to be independently re-verified by a human, unlike LLM-based auditing tools. Existing reproducibility infrastructure assumes the manuscript can be rebuilt from source. This tool addresses the opposite case: an artifact that is already published and can no longer be changed. Verification is not limited to quotations. Different disciplines admit different deterministic checks against a printed page, and these can be declared too: recomputing p from a printed test statistic and its degrees of freedom (t, F, chi-square, r — the same idea as statcheck, implemented independently, with the distributions validated against published statistical tables); the GRIM test for whether a reported mean of integer data is attainable at the reported sample size; check digits for ORCID, ISBN and ISSN; arithmetic relations among printed numbers such as randomised = analysed + lost to follow-up; and the correspondence between Japanese era years and Gregorian years. A claim that a printed value is wrong can itself be declared and checked. The tool's own test suite works by constructing a passing state and then breaking it in 26 distinct ways, asserting that each break causes exactly the corresponding check to fail. A checking tool cannot be trusted merely because it passes. AI use disclosure. The implementation of this tool was written using Claude Code (Anthropic). What to check, the deterministic-only policy, the declaration format, and the break-it-and-see-it-fail test design are the author's decisions. The AI is not an author; the author is responsible for all claims. DOI が付いて公開された成果物は改訂できない。改訂できるのは正誤表のほうであり、したがって正誤表は時間とともに一次資料からずれていく。引用が一字変わり、箇所が数え落とされ、未解決の項目が解決済みに書き換わり、最終更新の日付が古いまま残る。この道具は、そのずれを継続的インテグレーションで落とす。 判定に推論を使わない。大規模言語モデルも類似度も用いず、一致するかしないかだけで決めるため、出力を人が確かめ直す必要がない。既存の再現性の道具は原稿を作り直せることを前提とするが、この道具は作り直せない成果物を対象とする。 道具自身の検査は、通る状態を一つずつ壊して、壊した箇所がちょうど落ちることを確かめる形で書かれている。検査の道具は、通ることでは信用できない。