Imagine that two files arrive in a release folder with names that cannot be trusted. One was exported with Japanese publication settings and the other with English settings, but a handoff erased the useful labels. The safe response is not to guess from the filenames or from a Japanese sentence on the first page. Classify each stored package by its internal signature: the OPF language, navigation heading, colophon heading, and generated date.
The retained STUDIO-417 pair makes that incident reproducible. Both EPUBs came from the same four Japanese chapters. One package identifies itself as ja / もくじ / 奥付 / 2026年9月3日; the other identifies itself as en / Table of Contents / Colophon / September 3, 2026. This article treats EPUB language as a release-routing problem: determine which file belongs in which channel, prove that the pair still contains the same chapters, and send only the reader-specific question to a later test.
Quarantine Two Untrusted Build Names
Receive the files as package A and package B. Do not copy ja, en, “Japanese,” or “English” from their current names into the evidence sheet, because the purpose of the investigation is to discover whether those names are right. Record a SHA-256 digest for each original and place both in a read-only evidence location before changing any label.
In the fixed fixture, the Japanese-settings package has SHA-256 d6d3e66a…d46638, and the English-settings package has 6659aec3…73f2b1. Both were reopened as ZIP archives. Rune's local inspection returned valid=true for each, and the manifest check found zero missing resources. Those results establish that there are two inspectable packages. They do not yet say which is which.
The input identity is also known. It consists of four Japanese files titled as STUDIO-417 test chapters 1 through 4. Every chapter contains the sentence identifying it as a synthetic STUDIO-417 manuscript. Preserving the source description matters because a later chapter comparison will use that fixed set to detect a misdelivery, not to judge translation quality.
Let the OPF Assign Each Package
Open the OPF inside package A, then package B, and read dc:language. The fixture produces one ja value and one en value. That is the first reliable routing decision: assign the ja file to the Japanese-settings lane and the en file to the English-settings lane. The assignment comes from stored publication metadata, not from a release note or a visible chapter title.
Keep the field name precise. dc:language describes the publication's declared language. It does not report that a translation service processed the prose. A package can therefore carry en while the author-supplied test sentence remains Japanese. The evidence sheet needs separate cells for “package language” and “manuscript language” so that one cannot be mistaken for the other.
If both OPFs say ja, both say en, or one value is absent, stop the routing decision. Renaming a file would hide the failure. Retain the failed builds, return to the same four source files and their order, select the intended setting, and export to a new path. Reopen that candidate and read its OPF again before it receives a language label.
Fingerprint the Generated Pages With Three Strings
Metadata alone is not enough to release a file. Read the navigation and colophon XHTML and collect three generated strings. The ja package contains the navigation heading “もくじ,” the final-page heading “奥付,” and the date “2026年9月3日.” The en package contains “Table of Contents,” “Colophon,” and “September 3, 2026.”
The English colophon also contains “eBook,” “published,” “Author,” and “Publisher”; the Japanese counterpart contains “電子書籍,” “発行,” “作者,” and “発行所.” These values were read from the retained XHTML. An export-success response would not be a substitute, because the release decision concerns the bytes that actually arrived.
Treat the strings as a locale fingerprint. For example, an OPF containing en paired with “もくじ” is not an English package that merely needs a better filename. It is an internally inconsistent candidate. Move it to the failed lane, re-export from the fixed input, and recheck both the OPF and the three-string fingerprint. This makes inconsistency visible before a store or reader receives the file.
Match Four Chapters Before Routing the Pair
Correct labels cannot compensate for the wrong book. Both retained packages have four entries in the navigation document and four in the NCX. Their chapter targets run from test chapter 1 through test chapter 4 in the same order. Each spine has seven items: title, navigation, four chapters, and colophon. The missing-resource count is zero.
Put those values beside each other: chapter count 4, nav 4, NCX 4, spine 7, missing 0. The matching shape supports the conclusion that these are two derivatives of the controlled source set. It is also a practical misdelivery check. If package B has three chapters or a different target sequence, its en metadata does not make it the English-settings counterpart of package A.
Do not patch a missing chapter directly into the EPUB. Preserve the discrepant pair, compare the selected-file order with the four-source baseline, correct that input, and create a newly named build. Then repeat the entire stored-package inspection. A repaired archive is accepted only when its language fingerprint and its four-chapter structure agree with the intended lane.
Keep Manuscript Prose Out of the Locale Verdict
The English-settings chapter XHTML still contains the Japanese STUDIO-417 heading and Japanese synthetic sentence. For this fixture, that is an intentional control. The source was held constant so that changes in generated material could be attributed to the setting. The result shows that the setting reached the OPF and generated pages; it did not translate author-written text.
That distinction prevents two opposite mistakes. A production editor must not see “Table of Contents” and declare an English edition translated. A verifier must not see Japanese prose in this controlled English-settings package and replace it, because doing so would destroy the common-input condition used by the test. A real English edition requires separately translated and edited chapter files.
The handoff sheet should therefore label generated territory and author territory. OPF, navigation heading, colophon heading, colophon labels, and date belong to the first. Chapter headings and narrative belong to the second. The file can be classified by the first group without making an editorial promise about the second.
Route Date Rendering to a Named Reader Test
The fixed date input was 2026-09-03. Stored colophon XHTML renders it as “2026年9月3日” in the Japanese-settings file and “September 3, 2026” in the English-settings file. Those are measured package strings and belong in the locale fingerprint.
No Apple Books, Kindle, or other graphical reader was operated for this run. A reader may have a separate library panel, sorting field, or locale-dependent metadata display. Consequently, the evidence does not prove that every application will show the English month-first form, or that a device with another region will preserve it.
Create a separate reader-test ticket containing the exact EPUB SHA, application and version, device region, and screen to observe. If a later tester reports an unexpected date, first determine whether it appears on the packaged colophon page or in a reader-owned interface. Rebuilding a correct EPUB to compensate for an external library view would mix two different layers and could introduce a new defect.
Place the Files in Separate Release Lanes
The Japanese lane accepts the file whose stored values are ja, “もくじ,” “奥付,” and “2026年9月3日,” with nav 4, NCX 4, spine 7, no missing resources, and valid=true. The English lane accepts en, “Table of Contents,” “Colophon,” and “September 3, 2026,” with the same structural values. Each lane records the full SHA so that a renamed file can still be identified.
This decision is narrower than “both editions are publishable.” The fixture's Japanese prose is useful for isolating generated output, but it is not proof of an edited English manuscript. Reader-side date presentation also remains outside the package inspection. Translation and device behavior leave as explicitly named follow-up work rather than hidden assumptions.
The final release note should be short enough to use: original digest, assigned lane, OPF value, three-string fingerprint, four-chapter structure, and unresolved reader test. A recipient can reproduce the classification directly from the EPUB instead of trusting an earlier operator's memory. The result unique to this incident is a recovered destination for each anonymous build, not a general promise that changing one setting localizes an entire book.
Close the Incident With a Reproducible Identity
The two files began as untrusted names. They end as internally identified artifacts. Their OPF values provide the primary assignment, generated navigation and colophon strings confirm that assignment, and the matching four-chapter structure proves that they form the intended pair. The Japanese manuscript remains a documented control, while date display in graphical readers remains a documented open test.
That division makes the handoff resilient. Someone who receives the files next week can compute the digests, reopen the packages, and reach the same routing decision without the original filenames. When a language-switch build is at risk of misdelivery, internal identity—not appearance, habit, or a folder label—is the reliable release gate.