A benchmark for reviewing PowerPoint changes, not just rendering them

A slide can look unchanged while its stored behavior changes. A DrawingML
click, hover, or mouse-over declaration can point somewhere new; a
macro-category action can be retargeted while its public count remains the
same; opaque macro or embedded-object bytes can change without a visible text
edit. Those are review surfaces, but a tool’s claim to notice them has been
difficult to compare or reproduce.

Presentation Change Assurance Benchmark (PCAB)
is a narrow, reproducible answer. Its 0.1.0 release supplies twelve
deterministic baseline/candidate PresentationML pairs, public target-free truth
manifests, a structural validator, a portable observation protocol, and a
scorer. It evaluates a static presentation-review gate—not slide generation,
visual comparison, or runtime behavior.

Compatibility update: PCAB 0.2.0

After the initial release, I tested the generated packages with the independent
python-pptx reader. It exposed a package-skeleton omission: the compact
fixtures lacked the OPC root office-document relationship a general
PresentationML consumer expects. PCAB 0.2.0
adds that standards-form relationship to every package, validates it directly,
and loads all 24 baseline/candidate packages through python-pptx in CI and
fresh-wheel verification.

This is a versioned fixture-schema 2 update, not a quiet rewrite of the
original data. The twelve public facts and cases remain the same, while v0.1.0
remains immutable. One independent reader loading the packages is useful
compatibility evidence; it is not a claim that every PowerPoint client, runtime
action, or rendering feature has been tested.

One stored boundary at a time

Every PCAB pair preserves visible DrawingML text and changes exactly one ZIP
member. The cases cover external hyperlink retargeting; macro, program,
external-file, external-presentation, OLE-verb, media, internal-navigation,
and reserved actions; a missing action relationship; opaque VBA-project bytes;
and opaque embedded-object bytes.

The targets use reserved example.invalid endpoints. The macro and embedded
members are deliberately inert marker bytes, not executable samples. The source
generator and pcab validate prove archive integrity, deterministic bytes, the
one-member package boundary, stable visible text, and bounded action/relationship
shape. Public truth excludes action strings, relationship IDs and targets,
macro names, slide text, payload bytes, and payload digests.

A score that stays honest about scope

Tools can keep their own reports. An adapter emits a small normalized document
with cases it analyzed, exact public facts it found, and an optional allow,
review, or block reference disposition. Unsupported and errored cases are
explicit. PCAB reports expected-fact recall, analysis coverage,
reference-policy agreement, and complete cases.

The oracle is deliberately targeted, not exhaustive. An observation outside a
case’s declared fact remains visible as unrecognized evidence; PCAB does not
call it a false positive or manufacture a precision metric from a partial
oracle.

An optional SlideFence 0.2.0
adapter invokes its public JSON CLI and accepts a PCAB fact only when safe native
inventories and a change category exactly match the pair. It completes all 12
PCAB 0.1.0 cases with matching reference dispositions. That is an integration
result, not a requirement to use SlideFence or a claim that a declaration runs.

Static means static

PCAB does not open PowerPoint, render a slide, resolve or fetch a relationship,
execute VBA, activate OLE, play media, parse opaque payload formats, verify
signatures, or infer a client, policy, security, or runtime outcome. It records
stored package state only.

python -m pip install \
  https://github.com/SybilGambleyyu/presentation-change-benchmark/releases/download/v0.2.0/presentation_change_benchmark-0.2.0-py3-none-any.whl

pcab validate
pcab observation-template --output observations.json
pcab score --observations observations.json --output score.json

PCAB 0.2.0 has a hosted Python 3.11–3.13 matrix, fresh wheel and
source-distribution checks, released-wheel verification, and a separate
SlideFence integration job. The fixture corpus is also mirrored as a public
Hugging Face dataset.
The canonical note, release, and protocol are at
https://sybilgambleyyu.github.io/posts/presentation-change-assurance-benchmark.html.