Live: https://dadadevelopment.github.io/skill.state-showcase/
Interactive page for the Skill.State study (Terminal-Bench 2.0, GLM-5.3-flash, 87 tasks x 4 conditions x 3 trials). Static, no framework, no server. Every number is recomputed in the browser from the 1044 raw run rows of skill.state-paper/data/e1-release.
Poster: poster.html, PDFs site/skill-state-poster-ru.pdf and site/skill-state-poster-en.pdf (A1 landscape).
python3 scripts/build_data.py
node tests/stats.test.mjs
python3 scripts/export_html.py
node tests/export.test.mjs
python3 scripts/render_poster.pysite/is the page (open through any static server).dist/skill-state-showcase.htmlis one self-contained file (data, CSS and JS inlined, no network). Send this one.scripts/render_poster.pyprintssite/poster.htmlto A1 PDFs through headless Chrome.scripts/build_data.py --repo <path>reads the paper repository and writessite/data.json.tests/stats.test.mjschecks the browser statistics against the tables in the paper repository (solve rates, Wilson intervals, exact McNemar, cost / peak / output / cache ratios, horizon split).
- Summary by condition with filters (trial, task length), URL state (
?lang=en&trial=1&horizon=long). - Charts: solve rate with 95% interval, how runs ended (including
output_ceiling), cost of memory on a log scale. - Per-task heatmap, sortable.
- Runtime-state configurations from
data/followup-final/arm-summary.csvas an aggregate-only table. - A limitations section stating what is lost and what is not independently verified.
The page layout idea follows MTGBench (github.com/skyberg11/MTGBench, MIT). No MTGBench code, data or logos are copied; its licence is kept in licenses/ as the acknowledgement. Paper data licence: see the paper repository.