A production-ready benchmarking workspace for comparing LLM capability outputs under standardized task constraints.

每个模型生成的结果都放到这个路径：

public/submissions/<theme>/<model>/<submission-file>

例如：

public/submissions/clock/gemini-3.1-pro-high/index.html
public/submissions/clock/gpt-5.3-codex/index.html
public/submissions/carwash-decision/gemini-3.1-pro-high/response.md

放好后会自动被页面读取并渲染（无需改前端代码）：

• 本地开发：刷新 http://localhost:3000 即可看到
• Cloudflare 线上：提交并重新部署后可见

硬性要求（按主题类型）：

• 视觉主题（clock / recorder / weather-card / stock-panel / click-fireworks / neon-countdown / particle-gravity / cheetah-trophy-run / dslr-camera / schwarzschild-black-hole）文件名必须是 index.html
• 问答主题（carwash-decision）文件名推荐 response.md（也支持 answer.md / response.txt / answer.txt）
• 目录名必须是已支持主题之一（clock / recorder / weather-card / stock-panel / click-fireworks / neon-countdown / particle-gravity / cheetah-trophy-run / dslr-camera / schwarzschild-black-hole / carwash-decision）
• 一个模型一个目录（目录名就是模型名）
• 反作弊：严禁查看 public/submissions/<theme>/ 下其他模型目录；只能写入当前目标模型目录

When comparing model outputs, most evaluations are noisy because prompts, runtime, and rendering conditions vary. This project enforces a consistent benchmark surface so you can compare models fairly:

• Same task definitions
• Same output contract
• Same runtime container
• Side-by-side visual comparison in one page

• Unified benchmark dashboard (single-page comparison)
• Theme switching (Clock, Recorder, Weather Card, Stock Panel, Click Fireworks, Neon Countdown, Particle Gravity, Cheetah Trophy Run, DSLR Camera, Schwarzschild Black Hole, Carwash Decision)
• Model filtering (multi-select)
• Automatic submission discovery from filesystem
• Constraint inspection per submission:
  • total lines
  • CSS lines
  • JS lines
  • line-limit pass/fail
• Reusable prompt templates with hard constraints
• One-command prompt generator
• OpenNext + Cloudflare Workers deployment

• Next.js (App Router)
• React + TypeScript
• Filesystem-based submission ingestion (public/submissions/**)
• OpenNext for Cloudflare Workers

model-ui-arena/
  app/
    api/submissions/route.ts      # API endpoint: scan + aggregate submissions
    globals.css
    layout.tsx
    page.tsx
  components/
    arena-dashboard.tsx           # comparison UI
  lib/
    submissions.ts                # scanner + metrics + theme metadata
  prompts/
    base.md                       # shared hard constraints
    base-svg.md                   # unlimited inline SVG task constraints
    base-replica.md               # unlimited hand-drawn replica task constraints
    base-webgl.md                 # unlimited physics WebGL task constraints
    full-prompts.md               # full ready-to-use prompts
    themes/
      clock.md
      recorder.md
      weather-card.md
      stock-panel.md
      click-fireworks.md
      neon-countdown.md
      particle-gravity.md
      cheetah-trophy-run.md
      dslr-camera.md
      schwarzschild-black-hole.md
      carwash-decision.md
  public/
    submissions/
      <theme>/<model>/<submission-file> # model outputs (html or text)
  scripts/
    build-prompt.mjs              # prompt composer CLI

npm install
npm run dev

Open http://localhost:3000.

Every model output must be stored as:

public/submissions/<theme>/<model>/<submission-file>

Example:

public/submissions/
  clock/
    gpt-5.3-codex/index.html
    gemini-3.1-pro-high/index.html
  carwash-decision/
    gpt-5.3-codex/response.md
    gemini-3.1-pro-high/response.md

The dashboard auto-loads and renders all discovered submissions.

Anti-cheating policy:

• Never read or inspect other model directories under public/submissions/<theme>/.
• Only write to the current target model directory.

• clock
• recorder
• weather-card
• stock-panel
• click-fireworks
• neon-countdown
• particle-gravity
• cheetah-trophy-run
• dslr-camera
• schwarzschild-black-hole
• carwash-decision

npm run prompt -- --theme clock --model gpt-5.3-codex --max-lines 180 --language "HTML + CSS + JavaScript"

Use the generated prompt text directly in your target model.

Save the returned file into the matching submission path:

• visual themes: index.html
• carwash decision: response.md (recommended)

The benchmark page updates from filesystem scans.

• Runtime: visual themes use HTML + CSS + JavaScript; carwash-decision is text reasoning
• File count: exactly 1 per model per theme
• Max lines: 180-220 for standard visual themes; cheetah-trophy-run, dslr-camera and schwarzschild-black-hole have no code line limit
• External dependencies: forbidden
• Mobile baseline: width 390px

Returns:

• benchmark metadata (themes, constraints)
• normalized submission list
• per-submission metrics and pass/fail state

1. Run npx wrangler login once and confirm with npx wrangler whoami
2. Push repository changes to GitHub
3. Run npm run deploy
4. Verify model-ui-arena.lank.workers.dev

The deploy script builds the submissions manifest, creates the OpenNext bundle, uploads static assets, and deploys the Worker.

If you see hydration mismatch with injected attributes (e.g. theme/class attributes), this is usually caused by browser extensions mutating DOM before React hydration.

• Try Incognito mode (without extensions)
• Hard refresh (Cmd + Shift + R)

The layout already includes hydration-warning suppression on root nodes to reduce noisy extension-induced warnings.

If this appears during development, it is usually a build-cache collision (for example, running next dev and next build against the same .next directory).

• This project isolates dev artifacts to .next-dev (npm run dev)
• Production build artifacts remain in .next (npm run build / npm run start)
• If needed, clean both and restart:
  • rm -rf .next .next-dev
  • npm run dev

npm run dev
npm run build
npm run start
npm run prompt -- --theme <theme> --model <name> --max-lines <n>
npm run deploy

This project is licensed under the MIT License. See LICENSE for details.