The Calibration Tournament
§ Rules & format

The whole game.

00 ◊

Week 0 — the dry run

Week 0 is the practice week. It drops Sunday May 31, locks Monday June 1 EOD, and resolves Sunday June 7. Identical format to every other week, but scores from Week 0 do not count toward the official 10-week season.

Run it like the real thing. The cost of fumbling the interface, forgetting to lock, or submitting 99% on something you shouldn’t is zero — but the lesson sticks for the actual season.

Week 1 — the season opener — drops Sunday June 7. That’s where the scoring starts.

01 ◊

The shape

Ten weeks. Every week: ten yes/no questions about real-world events that resolve within seven days. For each one, you submit a probability between 1 and 99 — your honest confidence.

02 ◊

Scoring (Brier)

For each question, your score is (your_prob/100 − outcome)², where outcome is 1 if it happened, 0 if it didn’t. Lower is better. Your weekly score is the sum of 10 questions. Your season score is the sum across all locked-and-resolved weeks.

The math punishes overconfidence: saying 90% when the truth was roughly 60% hurts more than just saying 60% in the first place. That’s the whole pedagogy.

03 ◊

The cycle

Questions drop Sunday. Submissions lock Monday EOD. Outcomes resolve the following Sunday night. Leaderboard drops with the resolution. Next week’s set drops right after.

04 ◊

Anti-degenerate rules

  • 1–99 only. No 0%, no 100%. Real life never quite is.
  • Lock is final. One submission per (player, question) pair, enforced at the database level. No edit endpoint exists.
  • One forfeit per season. Miss a week, you score 0.25 per question — neither catastrophic nor free.
  • Question setter rotates.Whoever writes the week’s slate has an edge if they also play. So we trade weeks.
05 ◊

House bots

Luna and Claws play every week as calibration baselines. Their picks come in via the same database, but they’re flagged in the UI with a robot dingbat and are not prize-eligible. They exist to show you what raw AI predicts without the research process the rest of us are using.

06 ◊

Stakes

No money. The leaderboard winner gets naming rights for next season + a “Calibrated” tag in the channel. Dead-last faces a group-voted consequence (video roast, dinner, photo proof, group’s choice — voted before Week 1).

Separate “Most Improved” track: best second-half vs first-half Brier delta wins co-naming rights + forfeit immunity even if technically dead-last. Anyone can play for it.

07 ◊

Using AI

Allowed. Encouraged. Use any of Claude, ChatGPT, Gemini, Grok, Perplexity, Copilot, Meta AI, DeepSeek. Treat them as research tools, not oracles. Submit your own number. The gap between what AI says and what you submit is your judgment showing up — that’s the skill being trained.

← Back to the board
est. 2026 · ten weeks · ten questions

“The number you submit is yours. The math is just keeping score.”

Calibrate or perish.
✦ ◊ §
© 2026 The Practice · Brier-graded since day onev0.1.0
✦ ◊ § ✦ ◊ § ✦ ◊ § ✦ ◊ § ✦ ◊ § ✦ ◊ § ✦ ◊ § ✦ ◊ § ✦ ◊ § ✦ ◊ § ✦ ◊ § ✦ ◊ § ✦ ◊ § ✦ ◊ § ✦ ◊ § ✦ ◊ § ✦ ◊ § ✦ ◊ § ✦ ◊ § ✦ ◊ § ✦ ◊ § ✦ ◊ § ✦ ◊ § ✦ ◊ § ✦ ◊ § ✦ ◊ § ✦ ◊ § ✦ ◊ § ✦ ◊ § ✦ ◊ § ✦ ◊ § ✦ ◊ § ✦ ◊ § ✦ ◊ § ✦ ◊ § ✦ ◊ § ✦ ◊ § ✦ ◊ § ✦ ◊ § ✦ ◊ §✦ ◊ § ✦ ◊ § ✦ ◊ § ✦ ◊ § ✦ ◊ § ✦ ◊ § ✦ ◊ § ✦ ◊ § ✦ ◊ § ✦ ◊ § ✦ ◊ § ✦ ◊ § ✦ ◊ § ✦ ◊ § ✦ ◊ § ✦ ◊ § ✦ ◊ § ✦ ◊ § ✦ ◊ § ✦ ◊ § ✦ ◊ § ✦ ◊ § ✦ ◊ § ✦ ◊ § ✦ ◊ § ✦ ◊ § ✦ ◊ § ✦ ◊ § ✦ ◊ § ✦ ◊ § ✦ ◊ § ✦ ◊ § ✦ ◊ § ✦ ◊ § ✦ ◊ § ✦ ◊ § ✦ ◊ § ✦ ◊ § ✦ ◊ § ✦ ◊ §