Prediction MarketsCopy Anchor Link

Bet play-money points on who wins a debate, who wins a tournament, and whose team rating goes up. Every account starts with 1,000 points. Points are play money: they can't be bought, sold or cashed out, and they have no value outside the site.

• Route: /practice/predictions.
• Package: debate-predictions (pricing engine, settlement rules, validation, wire types, browser client, D1 migration), debate-webview (the page), and apps/debate-ai.com (API, queries).
• API: GET /api/predictions; GET /api/predictions/sources; POST /api/predictions/markets; POST /api/predictions/markets/[marketId]/bets; POST /api/predictions/markets/[marketId]/resolve

What it showsCopy Anchor Link

• Your points. Your balance. The 1,000 starting points land the first time you open the page while signed in. Guest accounts can watch but get no points.
• Top 5 teams. Pick a division (Public Forum, Lincoln-Douglas, Policy, College Policy) and bet on whether each of its five highest-rated teams gains team rating by the end of the month. The site opens these itself, a fresh set each month, and they settle themselves at month end.
• Major tournaments. For the same division, a winner market for each of the season's fifteen biggest tournaments still ahead, in date order: Heart of Texas, Florida Blue Key, Apple Valley, Glenbrooks, Princeton, Blake, Harvard-Westlake, Emory, Stanford, Harvard, Cal, the NDT (College Policy only), NDCA, the TOC and NSDA Nationals. Each lists the division's eight top-rated teams plus "Anyone else". Betting closes when the tournament starts, and a moderator settles it from the published results. Dates not yet published are the tournament's usual weekend, marked "expected dates".
• More open markets: the ones members opened, closing soonest first, filterable by type. Each market lists its outcomes with the current price as a percentage (the market's implied odds), how many points are staked, how it settles, and your position. The bet form quotes how many shares a stake buys, what it pays if that outcome wins, and where it moves the price.
• Open a market. Three kinds:
  • Debate: who wins one round. Type the two sides in, or pick an undecided round from a tournament hosted on the site.
  • Tournament winner: type the field in (up to 32 entries), or pick an event hosted on the site and every entry becomes an outcome.
  • Rating move: pick a team from the team rankings. The market asks whether its rating is higher at close than it is now.
• Settled: recent results, with what you got back.
• Top forecasters: the ten biggest balances, shown as ranks and points only (no names; your own row reads "You").

How it worksCopy Anchor Link

Pricing. Each market has an automated market maker using the logarithmic market scoring rule (LMSR), so a bet never waits for someone to take the other side. You stake whole points and receive shares; each share of the winning outcome pays 1 point. Buying an outcome raises its price. The engine is debate-predictions/src/lmsr.ts, which solves for the shares a stake buys in closed form and works in log space so large balances never overflow. The liquidity is DEFAULT_LIQUIDITY (150). The most a market can pay out beyond what was staked is b · ln(n) for n outcomes, which bounds how many points any one market can create.

Settling. Payouts are rounded down once per person, not per bet. Voiding a market refunds every stake. A market settles one of three ways:

• Hosted round (tabroom-panel): every judge on the round casts one vote, the entry their ballot's winloss score marks. Once all judges have voted, the majority wins. A bye, a forfeit or an even split voids it. Checked whether or not betting has closed, so a result that is in stops the betting.
• Hosted event (tabroom-event): the first-ranked entry in a published final result set (a bracket set, or one labelled final, champion, places or elim). Prelim seeds don't count.
• Rating (rating): once betting closes, the team's current rating is compared to the rating recorded when the market opened. Higher resolves "rises"; equal or lower resolves "doesn't rise". A team that has left the rankings voids it.
• Typed-in markets (manual) are settled by hand: by a moderator or admin, or by the creator while they hold no position in that market. Moderators can also settle or void any linked market whose data never arrives.

Hosted rounds and events are the debate-tournaments Tabroom tables in the same D1 database. Ratings come from the debate-rankings CSVs bundled with the build, read through debate-rankings-adapter on the site's scale.

Preset markets. debate-predictions/src/presets.ts holds the tournament calendar (MAJOR_TOURNAMENTS) and planPresetMarkets, which turns the rankings and the date into the markets that should be open. Their ids are stable (preset:rating:<dataset>:<hash>:<YYYY-MM>, preset:major:<slug>:<dataset>), so the board read (apps/debate-ai.com/lib/predictions/presets.ts) inserts only the missing ones and never reopens one that settled. They have no creator, so only moderators settle the tournament ones. A rating market settles "doesn't rise" when no rankings refresh ships during its month. Next season's calendar is a new MAJOR_TOURNAMENTS list.

Auto-settlement runs on every board read (resolveDueMarkets in apps/debate-ai.com/lib/predictions/queries.ts), up to 20 markets at a time. Every bettor gets a prediction_market notification when a market settles or is voided.

Guarded writes. D1 has no interactive transactions, so a bet is three guarded statements. First the wallet is debited only if it holds the stake. Then the market's share vector is written only if its version is still the one the price was computed from. If another bet got there first, the stake is refunded and the bettor is asked to retry at the new price. Finally the bet is recorded. Settling claims the market (status = 'open') before paying, so a market pays out once even if a moderator and the auto-resolver race.

Limits. Accounts need a real (non-guest) login to bet or open a market. Each account can have ten markets open at once. Betting closes between five minutes and a year after a market opens.

Tables. prediction_wallets has one row per account (balance and when the grant landed). prediction_markets has one row per market, with outcomes, share vector and source stored as JSON. prediction_bets has one row per bet, and its payout is filled in at settlement. They are declared in the app's schema.ts and shipped as packages/debate-predictions/migrations/, which the deploy's migrate-d1.ts applies.

Known gapsCopy Anchor Link

• You can't sell a position back before the market settles.
• Markets on rounds and tournaments that aren't hosted on the site (live Tabroom, round videos) must be typed in and settled by hand.
• Rankings only change when a deploy ships refreshed CSVs, so a rating market can only move when the site is redeployed with new rankings.
• Auto-settlement waits for someone to open the markets page; there's no scheduled job.
• A wallet never refills. Someone who loses every point can't bet again.

Markets the site opensCopy Anchor Link

Every Monday the site opens a rating market ("Will … gain rating?") on the top five teams in the Lincoln-Douglas and Public Forum rankings, each settling after 28 days. It also opens "Who wins …?" markets for the season's biggest tournaments (Florida Blue Key, Glenbrooks, Barkley Forum, Harvard and Cal), listing the current top eight plus "Someone else". Betting closes when the tournament starts and a moderator settles it from the final results. A team or tournament that already has an open market isn't opened twice.