Open technical preview
A screenshot is not a PowerPoint.
Gloss measures whether generated decks look right and are built right — native shapes, charts, tables, layouts, text, and relationships included.
Generative Layout & OOXML Scoring System · Apache-2.0
108 visual checks inspect what reaches the canvas.
- ▾ppt/
- ▾slides/slide relationships
- slide1.xml native shapes
- presentation.xml slide order
- ▾slideMasters/layout reuse
- slideMaster1.xml theme + fields
- ▾charts/editable data
- chart1.xml not a picture
172 package checks inspect what screenshots cannot.
The 21-second version
The benchmark begins where the screenshot ends.
The film is rendered from the same committed JSON bundle as this page. These are harness counts, not model scores.
One file. Two truths.
Visual fidelity is necessary. Native structure is non-negotiable.
A generated deck can match the reference image and still collapse the moment someone edits a chart, applies a layout, changes a theme, or reads the file with another tool. Gloss grades both layers.
Read the open scoring contractCandidate checks by method. Counts come from preview-v1.json.
The repository is the product
Find a weak assumption. Open a patch.
No waitlist and no private benchmark committee. Cases, assertions, evidence, and release gates are inspectable on GitHub.
Break the corpus
Add the slide construct your generator gets wrong: charts, tables, masters, fields, typography, geometry, or package relationships.
Propose a benchmark case ↗ CODEMake the grader stricter
Reproduce a false positive or false negative, then improve the assertion with a focused fixture and test.
Report a grader gap ↗ PROOFChallenge the evidence
Question a provenance record, mutation operator, scoring assumption, or release gate. Every public claim should survive review.
Challenge a claim ↗Truth before leaderboard
Working harness. Unfrozen benchmark.
Gloss is ready for public collaboration, not official model ranking. The release validator fails closed until independent evidence and frozen artifacts exist.
280 schema-valid checks and 381 traceable prompt clauses.
280 controls pass; 280 generated single-fault mutations are detected.
Assertion provenance and evidence remain incomplete for 280 candidate checks.
Prompt convergence, reviewer approvals, baselines, manifests, and signed release indexes remain open.
Reproduce it locally
Trust the command, not the claim.
git clone https://github.com/aronchick/gloss.git
cd gloss/acidslide-v1/grader
uv sync --extra dev --locked
uv run python ../benchmark/validate_corpus.py
uv run pytest tests/test_mutation_fixtures.py -qGloss is built in public