Skip to content

Repository files navigation

One Prompt — many models, no winner (so far)

A series of reproducible experiments: the identical prompt goes to several AI models at once — through their own CLIs, recorded live, with a correctness check baked into the deliverable. Whatever ships on the first pass is the result, published here running, not screenshotted, bills included.

The finding so far: for well-specified, checkable tasks, the model doesn't matter — the bill does. Every new test probes where that claim holds and where it breaks.

The tests

# Test Date Models Result
03 Double pendulum 2026-07-15 Sonnet 5 · Grok 4.5 · DeepSeek V4 Pro · Composer 2.5 Physics converged a third time (energy ≈6.6e-8% drift, positive λ) — the split was speed vs. self-verification: the two that ran a real browser took ~5× longer and were the only two you could trust
02 Lunar Lander 2026-07-11 Fable 5 · Grok Build · DeepSeek V4 Pro · Gemini→Gemma 4 Physics converged again (≈0% drift) — the autopilot split the field 20/20 · 18/20 · 5/20 · 0/20; the two that tested their own work landed
01 Fourier epicycles 2026-07-09 Fable 5 · Grok Build · DeepSeek V4 Pro · Gemini 3.5 Flash All four passed the referee (MSE ≈ 0.040, one rounding error apart); the bill spread ~200×

Every test ships: an overview (verdict, prompt, scoreboard), a live arena (all outputs running side by side), a blind test (shuffled + unlabeled — guess the author), one page per model (live demo + token receipts), and one page per measurement (accuracy / cost / size / speed).

Monorepo layout

packages/ui      @one-prompt/ui    Lit web components (op-*) + the one global stylesheet
packages/data    @one-prompt/data  typed test registry — one TS module per test
apps/site        @one-prompt/site  Vite multi-page site: home + one folder per test
.github/workflows/deploy.yml       builds and deploys to GitHub Pages on push to main
npm install
npm run dev       # local dev server
npm run build     # static build → apps/site/dist
npm run preview   # serve the build locally

The deployed site makes zero external requests — Lit is bundled, fonts are system, no CDN, no analytics. The model outputs in apps/site/public/tests/*/apps/ are byte-for-byte what the models wrote (never edited — they're the evidence).

Adding a test

  1. Data: packages/data/src/tests/<slug>.ts (a TestDefinition), prepend to TESTS in packages/data/src/index.ts. Home page, navs and pagers render from the registry automatically.
  2. Evidence: untouched outputs → apps/site/public/tests/<slug>/apps/, plus prompt.md and public/data/<slug>.json.
  3. Pages: copy apps/site/tests/fourier-epicycles/ as the reference implementation; register each page in apps/site/vite.config.ts.
  4. Follow DESIGN-SYSTEM.md — the contract for tokens, components, charts, and copy. The component kit is rendered live in the style guide.

Deploying

Push to main → the workflow builds and publishes. One-time setup: repo Settings → Pages → Source: GitHub Actions.

Provenance

  • Model identities were confirmed from each CLI's own session store (~/.claude/projects, ~/.grok, ~/.gemini, opencode's SQLite DB).
  • The site's hero animation is an independent implementation of the same epicycle math the models were tested on — it practices what it measures.
  • Write-ups and videos: @0xBakeer.

About

Reproducible experiments where one identical prompt goes to several AI models at once, through their own CLIs, recorded live with a correctness check built in. Whatever ships on the first pass is the result, published running rather than screenshotted, bills included.

Resources

Stars

1 star

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages