1,758 wallets hold a record of at least 20 resolved AI events, ranked by skill-adjusted accuracy across distinct events, so one correlated position can’t pass as a track record. The public board is free to read. The current #1 is 0x4b72…38d5, with 42 resolved events · +19.92¢ per share over the prices paid, after the sample-size adjustment · 75.0% hit rate at its Wilson lower bound. Its AI score, calibration diagram and graded history are free to read; the complete ranking behind it is the subscription. Scores are statistical estimates over public on-chain data, and they describe wallet records rather than persons. How the scoring works →
Free preview. The #1 forecaster’s identity and real numbers are shown. The rest are hidden, not blank. Pro sends alerts on Telegram and email, typically within minutes, when a tracked forecaster makes a move that clears the conviction gate, and shows every forecaster and score on the full ranking.Get the alerts and the full board →
The 5-minute price is the price of the first trade on the same outcome at least five minutes after each scored fill, whichever side started it (called the follower’s price on the research page). Each record in a category is re-priced there and measured before fees against the price curve, so the favorite-longshot premium anyone could get at that price is taken out. Held up means the average sat above zero at 95% confidence over at least 20 events at that price. Didn’t hold up means an average of zero or below over at least 20 such events, with no confidence test. Anything else has too little evidence for a verdict yet, including fewer than 20 events or a positive average short of 95%. The test treats a wallet’s events as independent, and related events can make a record look stronger than it is. It does not show whether buying at that price made money, and it is not the price a copier pays. In a pre-registered run on 2026-09-28, the taker price (the first purchase at the asking price, or for a sale the first sale at the bid, five or more minutes after each fill) cost 0.33 cents a share more on average, counting any price move in between. At that price, before fees, wallet-categories whose earlier records held up came out 0.21 cents a share below the price curve on later events, though about a cent a share better than the rest. Among records the board ranks above its baseline, the ones that didn’t hold up were not shown to do worse later than the rest (a pre-registered test on 2026-09-29). Labels as of September 27, 2026. The scores beside them are rebuilt nightly. A research label, not a recommendation to buy, sell, copy, or hold. How it was tested (the research page)
Scores are statistical estimates based on public on-chain data and may be incomplete or wrong. They describe a wallet’s historical record, and they are not statements about any person’s character, honesty, or overall trading ability. Wallet owner? Request delisting or a correction: hello@overround.pro.
So far 19 alerted positions on markets tagged AI have been graded: 13 won and 6 lost. Against the alert prices, the alerted side won 68.4% where the prices implied 57.0%, a difference of +11.5 points, on 19 resolved positions, too few for a 95% interval. Each position is listed, losses included, on the AI receipts page. Pro subscribers get alerts on Telegram and email, typically within minutes, from every category or only the ones they choose, for $19 a month at the founding rate.
AI markets price the industry's release calendar and milestones: whether a named model ships by a date, benchmark and capability claims, lab and company announcements. Resolution follows official releases and documented events.
Timelines are the graded skill. Release dates slip, announcements land early, and the market's price at entry states a probability about a specific deadline. Wallets ranked here have repeatedly judged those timelines better than the consensus across separate questions, through a period when the industry's pace made every deadline contested.
The Weekly Skill Report. Each Monday, one free email counts the positions the alert feed first flagged last week and how many have since won or lost. It names the busiest categories and the ledger’s running total, and links to every graded row.
Model release deadlines, capability and benchmark questions, and corporate events at AI labs. Markets resolve on official announcements or verifiable public evidence by the stated date.
How do deadline ladders grade?
Each date on the ladder is a separate yes-or-no market, so “released by March” and “released by June” count as two events even though one release settles both. Mutually exclusive sets, such as which company has the top model at month end, count once. Separate questions about different models, labs, and capabilities carry the most independent evidence.
Does following AI news closely produce a ranked record?
Reading the news is the baseline the market price already reflects. The graded question is whether entries beat those prices across at least 20 resolved events, which requires judging announcement timing and likelihood better than a consensus that reads the same feeds.