Sort Arena — Docs

Sort Arena

CADS-DEMO-sort (GitHub) is a CADS-Tunnel reference pipeline where nothing in the repo itself sorts anything. Every participant is a separate process, reached over the same real move contract, and the whole point is watching how differently the exact same underlying model behaves depending on what harness (system prompt, skill, algorithm coaching) is wrapped around it — not the prompt text.

The live arena is at sort.bunsenbrenner.org — a real, running visualization, not a static mockup.

Start here

New here? Five seconds to your first sorter. Clone, copy four commands, and a contract-verified participant is sorting — before any model call, so nothing can fail on you. Then one spec sentence makes it your own strategy, and the same page shows you why the harness, not the model, is what actually differs. Everything below goes deeper.

Want all of it on your own machine first? Run the arena locally brings up the same bridge locally in about a minute — every tutorial works against it unchanged.

If you’re here to learn the harness, not to sort

Read these six in order. They’re the argument, and each one is grounded in measurements from this arena rather than assertion. The fourth lets you check whether it landed, the fifth is the hardest variation, and the sixth is what happens when the measurements behind all of it are wrong:

  1. Five seconds to your first sorter — get something running, then see the same model produce measurably different participants.
  2. The harness is the variable — the four layers a participant can differ in, and why the contract decides which of an algorithm’s strengths can show up at all.
  3. Change the skill, not just the strategy — edit the driver. The worked example is a real defect where the contract already said the right thing and nothing checked, so it happened in 20 of 27 generations anyway.
  4. Check yourself — six predictions, each with a command that settles it. One of them catches a claim this site itself got wrong.
  5. Run it against your own model — the hard version, run to completion: swap where the model lives, change nothing else, and watch the same gates hold. A self-hosted model’s participant passes every check and finishes a real round in the hosted arena — proof the argument above was never a slogan.
  6. When the measurement lies — eight wrong claims published here, and what every one of them had in common. Two fail in opposite directions — one instrument missed real cases, the other invented them — and the last is the same question answered wrongly three times, each with more data than before. If you build gates on measurements, the measurement is load-bearing too.

Sections