export observatory

How well does export hold the pixels?

PPTX exports rendered as images and diffed against Gamma's own render — a frozen benchmark for the comparable accuracy number, rolling sweeps for what exists right now, and the corpus itself, deck by deck, card by card.

benchmark v1 · 29 decks · 403 slides · latest run 2026-08-20

frozen benchmark · pptx · 08/3195.97%benchmark v1 · 29 decks / 403 slides · weekly · comparable across runs
staging gammas · 11/0796.99%▼ 0.64pp~25 freshly sampled gammas · 2×/day · sample changes each run · 7-day mean 96.84%
official templates · 10/0196.92%▲ 0.05ppfull template catalog (633 decks) · Tue/Thu sweep
pdf export · 10/1498.31%▼ 0.04ppbenchmark v2 · 39 decks / 606 slides · Gamma PDF export rendered via poppler · same frozen corpus
png export · 08/2599.6%benchmark v2 · 39 decks / 606 slides · Gamma PNG export, near-truth pixels · pipeline canary · same frozen corpus
gslides export · 08/2797.21%benchmark v2 · 36 decks / 562 slides · Gamma PPTX export Drive-imported to Google Slides, rendered by Google · same frozen corpus

each number is visual fidelity, 100 − mean changed pixels per slide, from that series' latest run · Δ compares the previous run of the same series

← back to the observatory
01

Export fidelity over time

The canonical export accuracy number: visual fidelity of the PPTX exporter over the frozen, versioned export-decks benchmark — exported decks rendered as images and diffed against Gamma's own render, measured weekly.

02

Run statistics, week by week

The distribution behind each weekly headline: per-cohort mean, median, p90, p95, and worst-slide changedPct. The tail percentiles show whether a score moved because everything shifted or because a few slides broke.

pdf export · poppler render

cohortscopenfidelitymean changedp50 changedp90 changedp95 changedworst changedΔ mean
overallper slide60698.31%1.69%1.02%2.75%4.57%47.97%—
per deck39—1.61%1.32%2.91%4.33%4.68%-0.08pp
templatesper slide43998.28%1.72%0.96%2.5%5.22%47.97%—
per deck19—1.74%1.32%2.71%4.56%4.68%+0.02pp
importsper slide6399%1%0.89%1.6%2.03%4.44%—
per deck10—1.1%0.92%1.61%2.23%2.86%+0.1pp
remixesper slide10498.01%1.99%1.35%4.05%4.94%15.41%—
per deck10—1.87%1.75%3.23%3.77%4.31%-0.12pp

Per-slide rows pool every slide's changedPct in the cohort, so large decks weigh more. Per-deck rows take each deck's mean changedPct first, so every deck weighs the same. Δ mean is per-deck mean minus per-slide mean in percentage points: positive means small decks export worse than the pooled score suggests, negative means large decks do. Worst is the single most-changed slide (or deck). Runs before percentile capture show —. Runs that cover extra export formats get one table per format.

03

Rolling export sweeps

Export fidelity of what exists right now — the current official template catalog (Tue/Thu, full corpus) and freshly sampled staging gammas (twice daily), plotted as 100 − mean changedPct per run. The corpus shifts as templates and gammas change, so these lines show coverage and discovery; the frozen benchmark above is the comparable accuracy number.

04

Structural fidelity

What survives export as native OOXML, independent of pixels: chart and table survival, image drops, autofit health, and typed shape-level deltas from the struct-diff suites — ideal-GML → export over grid-templates, and the full PPTX round trip (source → import → export) with per-leg loss attribution.

tables native
34/37
91.9%
image drops
0
autofit missing fontScale
1673/1673
decks
34 run
6 failed · grid-templates · cross-kind
top structural deltas (latest run)
dropped · 5814resized · 5182moved · 5147font-size · 4385added · 2997
05

What the benchmark covers

The frozen corpus roster: cohort sizes, card counts, locales, scripts, and themes, with each deck's provenance. The roster changes only at deliberate benchmark version bumps, so scores stay comparable within a version.

40 decks630 cards20 template · 463 cards10 import · 63 cards10 remix · 104 cardslocales beyond en: non-en, he, zh-CN, ja, pt-BR, ko, es, de3 dark-theme1 RTL
06

The corpus, deck by deck

Every deck in the frozen benchmark, sorted worst-first from the latest baseline run. Open a deck to compare each card three ways: Gamma's own render (truth), the render of the PPTX export, and the pixel-diff heatmap between them.

sorted worst-first · run 2026-08-20 · benchmark v2
07

Factory runs

Triage-and-fix sweeps over the ticket queues — export, templates, frontier, and import — with what each run fixed, routed to a human, or skipped as already claimed, plus the draft PRs it produced.

4 runs · 4 fixed · 9 routed to humans · 1 bounced · 31 skipped as in-progress · 6 PRs

datearearunfixedroutedbouncedskippedPRs
2026-09-04exportpptx-export fix loop: EDI-12269 painted custom-mask accents1 (EDI-12269)000gamma#35172
2026-08-31importimport-fidelity backlog triage + fix1 (EDI-11392)204gamma#34099 · import-decks-dataset#2202
2026-08-31frontierexpressiveness-factory queue run1 (EDI-7976)616gamma#34097
2026-08-31templatestemplate-library triage + fix sweep1 (EDI-11340)1021gamma#34096 · template-library#289

Fixed = tickets that produced a draft PR with evidence. Routed = tickets handed to a human with a diagnosis comment (product/design decisions, unsafe or blocked repros). Bounced = tickets returned for missing information. Skipped = tickets left untouched because another session, assignee, or open PR already had them.