REFERENCES & ATTRIBUTION

The work behind the references.

Reference screenshots retain the terms of their original sources. These selected frames were resized and encoded as WebP for the project website. The entries below preserve source information recorded in the released benchmark.

Task 280 · Air quality command center

Reference A for Task 280: Palette & atmosphere

Reference A · Palette & atmosphere

eventre.

Reference B for Task 280: Cards & controls

Reference B · Cards & controls

A synthetic reference webpage from the benchmark’s curated source pool.

Reference C for Task 280: Full-screen layout

Reference C · Full-screen layout

A synthetic reference webpage from the benchmark’s curated source pool.

Generated output: GPT-6 Astra · final artifact from a recorded OpenCode rerun. A replayed screenshot of the final artifact in the recorded air-quality case. This example is separate from the frozen leaderboard run.

Task 35 · A bilingual seismic observatory

Reference A for Task 35: Palette & components

Reference A · Palette & components

A synthetic reference webpage from the benchmark’s curated source pool.

Reference B for Task 35: Dense technical grid

Reference B · Dense technical grid

A synthetic reference webpage from the benchmark’s curated source pool.

Generated output: GPT-6 Astra · frozen final HTML, replayed locally. The generated page uses illustrative teaching data. The screenshot was replayed on 13 October 2026; the page’s clock and dates are dynamic.

Task 125 · Two Rivers clothing collection

Reference A for Task 125: Asymmetric layout & grid

Reference A · Asymmetric layout & grid

influencer · openly licensed template from the benchmark’s reference pool. Source terms are recorded in the structured attribution file below.

Reference B for Task 125: Dark palette & accents

Reference B · Dark palette & accents

A synthetic reference webpage from the benchmark’s curated source pool.

Reference C for Task 125: Avoid thin typography

Reference C · Avoid thin typography

small-apps · openly licensed template from the benchmark’s reference pool. Source terms are recorded in the structured attribution file below. Requested departure, not a match.

Reference D for Task 125: Rounded cards & pills

Reference D · Rounded cards & pills

A synthetic reference webpage from the benchmark’s curated source pool.

Reference E for Task 125: Avoid extreme type contrast

Reference E · Avoid extreme type contrast

A synthetic reference webpage from the benchmark’s curated source pool. Requested departure, not a match.

Generated output: GPT-6 Astra · frozen benchmark final artifact Screenshot replayed from the unmodified, score-matched frozen HTML.

Task 273 · Silent Radio mini-game arcade

Reference A for Task 273: Badges & form controls

Reference A · Badges & form controls

educenter · openly licensed template from the benchmark’s reference pool. Source terms are recorded in the structured attribution file below.

Reference B for Task 273: Compact grid & code block

Reference B · Compact grid & code block

A synthetic reference webpage from the benchmark’s curated source pool.

Reference C for Task 273: Bold typography

Reference C · Bold typography

A synthetic reference webpage from the benchmark’s curated source pool.

Generated output: GPT-6 Astra · frozen benchmark final artifact Screenshot replayed from the unmodified, score-matched frozen HTML.

Typography

DM Sans, Manrope, and Space Mono are distributed under the SIL Open Font License. Local copies keep this website independent of external font services.

DM Sans license · Manrope license · Space Mono license

DejaVu Math TeX Gyre provides the local mathematical typesetting. DejaVu Math license

Website & research

The website design and implementation are original. Paper content, example selection, and results follow the RefCraft manuscript and repository. Model names identify the evaluated systems.

The project mascot and scientific figure panels are extracted from the supplied RefCraft manuscript. The mascot retains its original transparency; paper figures retain their original colors, labels, and data.

Download structured reference attribution

Model marks

The nine model marks shown in the leaderboard and the performance–token plot are rendered directly from Table 1 of the supplied RefCraft manuscript, at their original artwork and colors. Each mark’s page rectangle, extraction settings, and SHA-256 are recorded in data/model-marks.json; scripts/extract_model_marks.py reproduces them. The marks identify the evaluated systems and imply no endorsement by their owners.

Official project marks

Qwen · Official Qwen blog logo. The mark retains its original artwork and colors.

GitHub · Official GitHub Mark asset. The mark retains its original artwork and colors.

Hugging Face · Official Hugging Face organization avatar. The mark retains its original artwork and colors.