Release note · August 4, 2026
Stored but unbound: custom XML is a document review surface
A document-review benchmark should not require a visible content control before it can test a stored custom-XML change. A package can carry a conventional custom-XML store without a mapping, and a static reviewer should be able to account for that storage boundary without inventing client-runtime behavior.
Document Change Assurance Benchmark (DCAB) 0.23.0 adds its thirty-fourth deterministic pair: review.unbound_custom_xml_payload_changed. The pair changes only a synthetic custom-XML payload. It is deliberately not a data-binding case.
One opaque boundary, stable package topology
Microsoft’s Custom XML parts overview describes embedded custom XML parts as package data that Office solutions can create or modify while a document is open or closed. That makes stored presence independently reviewable; it does not establish that a particular document view will surface the data.
customXml/item1.xml only changed member
customXml/itemProps1.xml byte-identical
customXml/_rels/item1.xml.rels byte-identical
word/document.xml byte-identical
w:dataBinding absent on both sidesBoth files retain the same package-member set, conventional custom-XML data/properties topology, relationships, and sequence of stored w:t values. The builder changes only inert synthetic XML bytes in customXml/item1.xml. It does not parse that payload as application data, resolve a target, open Word, render a document, update a control, or claim that Word displays, uses, or removes it.
A target-free truth envelope
The public truth says only that the pair represents an unbound custom-XML payload boundary:
{
"kind": "unbound_custom_xml_payload_changed"
}It omits payload values, XML nodes, namespaces, identifiers, and any meaning assigned to the stored data. Structural validation checks the single-member boundary, the absence of a w:dataBinding marker, unchanged relationship and properties parts, stable stored text, ZIP and XML validity, independent reader compatibility, and deterministic regeneration.
One released consumer, still a tool-neutral corpus
The optional DocFence 0.33.0 adapter reaches a strict 34/34 score with public aggregate evidence only. For this pair it requires a generic custom-XML change, an unchanged two-part custom-XML inventory, and zero data bindings. It never receives the XML payload or a private fingerprint.
Hosted CI passed on Python 3.11, 3.12, and 3.13, including a fresh public-DocFence 0.33.0 adapter installation. Fresh wheel and source-distribution installs validate the bundled 34-case corpus. The public Hugging Face dataset mirror was atomically synchronized after release; its complete project-file inventory and the new pair’s hashes were compared with the tagged source.
Use the release
python -m pip install https://github.com/SybilGambleyyu/document-change-benchmark/releases/download/v0.23.0/document_change_benchmark-0.23.0-py3-none-any.whl
dcab validate
dcab docfence-observations --executable docfence --output observations
dcab score --observations observations --strictFixture schema version 1 stays stable because the public truth and observation envelopes are unchanged. DCAB 0.23 contributes one exact static-review fact, not a custom-XML classifier, a document scrubber, a Word client, a renderer, or a promise about application behavior. The MIT-licensed builder, fixtures, verifier, release artifacts, and research notes are available on GitHub.