The Archive That Disagrees With Itself

The useful question is not whether the first four bytes say PK. It is whether the directory at the end points to complete local records, whether the sizes and CRCs agree after the one explicit exception is applied, and whether the bytes decompress to what either record promised.

The instrument is already running on a document made for this page. Bring nothing and every panel still works. Bring a file and the fixture steps aside. The complete file is read once into one ArrayBuffer. It is never uploaded, extracted to disk, executed, rendered as markup, placed in storage, or offered to any relationship target.

The published anchor comes first

Three signatures, three fixed lengths, one exception

PKWARE APPNOTE 6.3.10 prints the local file header signature as 0x04034b50, the central file header signature as 0x02014b50, and End of Central Directory as 0x06054b50. Their fixed portions are respectively 30, 46, and 22 bytes, before their declared variable fields. Every multibyte integer used here is unsigned and little-endian.

Archive Fork A contains five original XML parts. Its word/document.xml entry sets general-purpose bit 3. The local header therefore stores CRC-32, compressed size, and uncompressed size as zero. The signed data descriptor after the stored bytes carries the values that agree with the central record. The three raw differences are an EXPECTED BIT 3 result, not a conflict.

This is the first check because it can fail in both directions. A checker that treats every unequal byte as corruption rejects a permitted package. A checker that simply trusts the directory accepts a contradictory package without noticing that the corresponding record said something else.

The central directory is not called universally authoritative here, because APPNOTE does not appoint a universal winner for contradictory copies. Its authority is emergent rather than decreed: the directory is simply the only guaranteed way to enumerate an archive, since files may be stored in any order, a data descriptor may be unsigned, and the archive may sit behind a self-extracting stub. What follows from that is a distinction between classes of reader, not a rule about all readers. A random-access reader enumerates through the directory. A streaming or front-walking reader never sees the directory at all and takes the local record's word for it. A careful reader notices the disagreement and says so. For a generic ZIP this instrument reports UNEXPLAINED DIFFERENCE and stops, and it does not predict which of those a named extractor is.

ECMA-376 supplies the narrower rule for an Open Packaging Conventions package: corresponding appropriate values must be equal after the bit 3 data-descriptor case is resolved. That is why an OPC package can receive an Annex B.2 pass or fail while an APK cannot receive that verdict from this page. It also means a conforming Office document cannot disagree with itself in these fields, because Annex B.2 forbids it. Self-disagreement is a property of generic ZIP tooling and of packages that violate that clause, not of valid .docx files. So an Annex B.2 pass on a file your word processor wrote is the expected outcome and is close to no news at all. The informative outcomes are the failures, the entries the directory never mentions, and the differences a generic archive is free to carry.

Layer two · the package above ZIP

The directory is only the floor

A dismissal is available: these are duplicated bookkeeping fields, and a bookkeeping viewer has learned nothing about the document. So the next instrument follows the package's own declarations. It recognizes OPC only when [Content_Types].xml and _rels/.rels exist, carry the required package namespaces, and were safely decoded. Then it resolves internal relationship targets, keeps external targets inert, audits name equivalence and compression restrictions, and locates Word parts through relationships rather than guessed filenames.

An OPC syntax pass does not prove that Word, Excel, PowerPoint, an EPUB reader, Java, Android, or any other named product will accept the package. Those profiles add rules this page does not implement. XML is shown only as escaped text. Macros, scripts, SVG, HTML, media, and external relationship targets are never executed or requested.

What an RSID can say, and what it cannot

ECMA-376 Part 1 section 17.15.1.70 says that identical revision identifiers in documents with the same rsidRoot indicate the same editing session. Archive Fork A and B use the synthetic root 10000001. They share session identifiers 10000001 and 10000002, then A alone carries 10000003 and B alone carries 10000004. The page displays those as unordered sets.

Spennemann and Spennemann used this signal in a 2023 proof of concept involving more than 400 related templates from one publisher. That work demonstrates a genealogy method in that corpus, not an error rate for arbitrary pairs. Spennemann and Singh reported in 2024 that Microsoft Word does not reliably follow the increasing-value rule. This instrument never sorts RSIDs into a chronology.