A linked picture target can change without changing Word text

A Word document can keep every stored text node, every DrawingML picture
marker, and every relationship ID fixed while changing the target of a linked
picture. That is a stored review boundary a text diff misses, but a static
reviewer can expose without downloading an image.

Document Change Assurance Benchmark 0.3.0
adds its fourteenth deterministic pair:
external.drawing_linked_picture_target_retargeted. Both sides contain the
same a:blip r:link marker and package members. Only
word/_rels/document.xml.rels changes, retargeting a standard external image
relationship.

Recognize the marker; do not retrieve the image

Microsoft’s Open XML documentation for Blip.Link
describes r:link as a linked-picture reference for an image outside the file.
DCAB models that relationship-backed marker, but never resolves a target, opens
a document client, renders an image, or claims that any Office client will load
it.

The package target is synthetic example.invalid data. Its public truth names
only drawing_linked_picture_target_changed, with an external binding and
image relationship category. It excludes the target, relationship ID,
relationship path, and presentation metadata.

A deterministic static boundary

The fixture uses a compact DrawingML inline picture with a:blip r:link and no
image payload member. The candidate changes only the relationship target. An
independent verifier checks the marker, absence of r:embed, standard image
relationship type, external target mode, deterministic bytes, stable member
set, and exact pair boundary.

The standard python-docx reader opens all 26 .docx fixtures, and its
lower-level OPC reader opens all 28 packages. The optional local
DocFence 0.27.0
adapter maps public aggregate external-relationship and linked-picture-inventory
evidence without consuming targets or private signatures.

DCAB 0.3.0 keeps fixture schema version 1 because the public envelopes are
unchanged; it extends the corpus from 13 to 14 cases. It does not claim image
safety, a network request, a renderer result, client behavior, or one universal
policy.

The source, generated corpus, and verifier are MIT-licensed on
GitHub. Read the
canonical release note
for the full boundary and install command.