A Word field can quietly consult the outside world

A field result can look like ordinary document text while its stored instruction
says something materially different: query a database, include another
document, pull a picture, retain an OLE link, refer to a document used by an
index, or exchange data with another application. A generic field count cannot
distinguish a harmless date from one of those outside-facing instructions.

DocFence 0.8.0
adds a deliberately bounded, privacy-safe inventory for that difference. It
recognizes DATABASE, legacy DATA, DDE, DDEAUTO,
INCLUDE/INCLUDETEXT, INCLUDEPICTURE/IMPORT, LINK, and RD fields,
then reports only category counts. Field instructions, source paths, connection
strings, queries, application names, item references, and private fingerprints
do not enter JSON, Markdown, or SARIF.

Why the field encoding matters

OOXML stores a simple field in w:fldSimple/@w:instr. A complex field is a
begin/separate/end sequence whose pre-separator instruction may be split across
multiple w:instrText runs and may contain nested fields. DocFence reconstructs
that local instruction state, classifies only its initial field keyword, hashes
the full instruction privately, and discards the source-bearing value before
public output is created.

Tracked revisions add a subtle failure mode. Word can preserve a deleted field
code in w:delInstrText beside a current w:instrText code. Concatenating them
creates an invented field and can hide a historic external reference. DocFence
assembles current and deleted variants separately, including moved-from and
moved-to field-code ranges. A revised field can therefore contribute evidence
for both stored forms without disclosing either instruction.

Evidence, not execution

The field name is a review signal, not a runtime verdict. DocFence does not
parse field arguments, run a database query, start DDE, open an OLE object,
locate an included file, retrieve content, or decide whether Word would update a
field. It does not claim to inventory every Word expression that can display a
URL. The boundary is deliberately limited to stored evidence for named field
families.

For a clean handoff, reject all recognized external-source field instructions:

rules:
  require_no_external_fields: true

That emits DFP027 when any category count is nonzero. For an approved template
with known fields, freeze a controlled baseline instead:

rules:
  no_external_field_changes: true

This emits DFP028 if the private inventory changes. Run fragmentation alone
stays quiet; a changed stored source does not.

Release evidence

The release covers all eight field categories, simple and complex encodings,
split, nested, resultless, Strict, header-story, revision, moved-field,
non-field-text, unclosed-field, policy, SARIF, and redaction cases. The final
wheel also profiled Apache POI’s independent
FieldCodes.docx
and FldSimple.docx
fixtures without a false external-source count.

The wheel and source archive were built independently twice with the same
SOURCE_DATE_EPOCH and matched byte-for-byte. The public release includes a
wheel, source archive, and SHA-256 manifest; I downloaded the published assets,
verified the checksums, installed the wheel in a fresh environment, and reran
the fixture profile.

Install it with:

python -m pip install https://github.com/SybilGambleyyu/docfence/releases/download/v0.8.0/docfence-0.8.0-py3-none-any.whl

The complete explanation, sources, policy scope, and limits are on the
canonical release note.