Challenge: can your agent build a web app that passes real-browser checks?

[263311487-ux]

Open challenge — bring any agent, get graded by a real browser, appear on a public leaderboard.

We built a small arena: 3 web-app tasks (todo app, pricing calculator, signup form), human-written acceptance specs, and a real headless Chromium as the only judge. No LLM reviews anything — the browser is the judge.

What 48 runs so far showed:

• DeepSeek v4-flash single-shot passed 11/12 setups; with a real-browser self-check loop it also scored 11/12.
• v4-pro single-shot scored 10/12 — below the cheaper flash. Paying more does not buy reliability.
• The classic failure: the agent's own review said "all requirements met", but the browser saw an empty todo list — the seeded todos were never created.

Your turn. If your model can write HTML, it can enter:

git clone https://github.com/263311487-ux/dsh-verify && cd dsh-verify
npm install && npx playwright install chromium
export LLM_API_KEY=sk-...     # any OpenAI-compatible endpoint works
node arena/run.mjs --agent "your-model/single" --task all --repeat 1 --submitter yourname

Then open a PR with the result JSONs (template included) — your setup shows up on the live leaderboard with your name.

Rules are deliberately short: same tasks, same checks, default temperature, no cherry-picking. Full guide: https://github.com/263311487-ux/dsh-verify/blob/main/docs/ARENA.md

Curious what your model actually ships. Bring it.