Classroom Mode

Two features turn a solo arena into a room activity: the thinking selector and the audience vote. Neither needs configuration.

Thinking selector

The toolbar under the prompt has a thinking: dropdown: off, low, medium, high. The value is sent to both models as reasoning_effort and recorded on the battle, the export, and the reveal.

• Models that stream a reasoning trace (reasoning_content) get a collapsible thinking block above the answer. It opens while the model thinks and collapses once the answer arrives, with a character and reasoning-token count in the summary.
• A model that rejects the parameter with a 400 is retried once without it. The panel header says “thinking not supported here, answering plainly” and the battle proceeds. The reveal shows the effort that was actually applied, which is off for that side.
• To skip the retry dance, pin the behavior per model in models.yaml:

  - id: deepseek-r1
    reasoning: true     # supports reasoning_effort; a 400 is a real error
  - id: llama3-8b
    reasoning: false    # never send reasoning_effort

Leaving reasoning out means auto, which is right for most gateways.

Pick one model, draw the other

The two model selectors are independent now. Leave both on “mystery match” for a blind draw, set both for a head-to-head, or set one and let the arena pick its opponent from the same category. The model you chose keeps the slot you put it in.

Audience vote

1. Run a battle and wait for both answers to finish.
2. Under the vote buttons, click let the audience vote (phones).
3. A panel opens with a QR code, the join URL, and a six-character code. Put it on the projector.
4. Students open the link on their phones. They see the prompt, both responses, and three buttons. Each phone gets one vote, changeable until the poll closes. The running count is visible; the split is not, so early votes cannot pull later ones.
5. Watch the live tally bars on your screen, then click CLOSE POLL & REVEAL. The plurality becomes the battle’s recorded vote with method = "audience", ELO updates as usual, and every phone flips to the reveal: which model was A, which was B, the tally, and whether their pick matched the room.

You can still vote by hand or hand the battle to the judge while a poll is open. Either one wins the race: closing the poll afterward returns 409 and the poll is simply abandoned.

What the phones can and cannot do

The audience surface (/vote/<code> and /api/audience/...) sits outside the passphrase gate on purpose. It can read the two finished responses for its poll and record one choice per device. It cannot create battles, reach a model provider, read the leaderboard, or see model names before the poll closes.

Guardrails: codes come from an unambiguous 31-character alphabet, polls expire six hours after opening, each poll caps at 1,000 distinct voters, and the vote endpoint is limited to 300 requests per minute per IP. That last number is deliberately loose: a whole class usually arrives from one NAT address or one tunnel, and the real cap is per-poll voters.

Remote classrooms

The join URL is whatever origin the presenter’s browser is on. If the arena is only reachable on a LAN, phones must be on that LAN. Put it behind a tunnel (see Deployment) and the QR code works from anywhere.