# Original PDF export experiment / Eigener PDF-Exportversuch

Wrapped.li, October 31, 2026. Automated original file experiment; no independent human reviewer or real-reader study. All content is constructed. No account or customer data, third-party artwork, LinkedIn upload or paid service was used.

Four original four-page 540pt square documents (brief, table, diagram, synthetic pattern), seven conditions each, all 28 PDFs retained. The 112 exported pages are repeated treatments of 16 original pages, NOT 112 independent designs or readers. Protocol registered before export. A second complete local run reproduced all PDF hashes and per-page metrics; no conditions were discarded.

## Reproduce

Use an existing Node workspace with pdf-lib 1.17.1 and sharp 0.35.5. Put reproduce.mjs where those modules resolve and make pdftoppm available on PATH. Run:

    node reproduce.mjs /path/to/new-output-directory

The output directory contains the 28 PDFs, results.json, results.csv and temporary page renders. It creates only files belonging to this experiment. Keep the public script and protocol with the result. The recorded run used Node 24.21.0 and Poppler pdftoppm 26.05.0. JSON includes detailed sharp dependencies. Base-14 Helvetica/HelveticaBold, fixed PDF dates, uncompressed object streams, PNG compression level 9, Lanczos3 resize and JPEG 4:4:4 are defined in the script. There is no stochastic input.

Each raster export begins with the vector PDF rendered by Poppler at 1080px, then uses the stated resolution and encoding before PDF embedding at unchanged page geometry. Pixel metrics use a second Poppler render at 360px and compare every RGB channel against the vector page rendered directly at 360px. RMSE = sqrt(mean((reference_channel - output_channel)^2)); mean document RMSE is the arithmetic average of its four page RMSEs. Values are on the channel scale 0–255. Different rendering pipelines, antialiasing and pixel alignment affect this number. It is NOT a readability, perceptual-similarity, accessibility or LinkedIn-processing score. The vector baseline is zero by definition.

Text-drawing-block counts inspect BT operators in decoded page content streams. Embedded-image counts inspect page XObjects. These are structural observations, NOT text-extraction accuracy, correct reading order, tags or screen-reader acceptance. The vector PDFs are untagged and are NOT accessibility-certified.

Use results.csv for all bytes and per-page metrics; results.json includes every file SHA-256, conditions and provenance. complete-study.zip includes all 28 PDFs plus generator, data, method and protocol. Standalone sample PDFs and graphics are included for convenient inspection. The preview graphic summarizes each document's full raster size range; the article table and data retain all seven conditions. The document artwork is the same English test material in both localized articles; translated discussion and graph labels are provided separately.

## Limits / Grenzen

Only these four constructed text/shape designs and this export pipeline were tested. There are no photographs, platform uploads, reader tasks, assisted-technology tests, OCR accuracy tests, causal inferences or population estimates. Do not infer an optimal LinkedIn size, algorithmic ranking, human preference or business lift. A repeat is reproducibility evidence, not a second independent sample. Pixel alignment and font antialiasing can produce large error without measured loss of meaning. Inspect the final publication destination separately.

The experiment source, constructed artwork and dataset are original Wrapped.li materials, published to make the method and results inspectable and reproducible. External pdf-lib, sharp and Poppler dependencies retain their own licenses and distribution requirements. No third-party code or executable is bundled.