Exhibit: the method
The method, in full
Published so you can disagree with it — and check it.
Every competitor in this category sells a proprietary score. Ours is written down here, in the same words the code runs on, with the public endpoints you need to compute it yourself. If the rule changes, the version number changes with it and the old calls keep the rule they were made under.
Exhibit A — the call rule, verbatim
Call rule v2.0 · in forceA call is emitted when all of the following are true:
- WHO — the wallet is in the tracked cohort at the moment of the trade. The cohort is the top 100 RATED wallets by composite score, rebuilt daily from the public trade tape. A wallet is RATED only with at least 8 settled trips across at least 4 distinct markets; it enters the cohort only with a positive settled edge (realized ROI > 0 on money actually risked) and a composite score in the top decile (≥ the 90th percentile) of every rated wallet scored in that day’s refresh.
- WHAT — the wallet OPENS A NEW position: a market + outcome it has no prior recorded position in. Adding to a position it already holds is not a call.
- HOW BIG — the new position is at least 3× the wallet’s own median settled stake, and at least $500. The multiple is the signal; the floor exists so a wallet whose median stake is a few dollars cannot emit calls on pocket change.
- WHEN — the market is open and its real decision clock is more than 24 hours and no more than 120 days away. Sports use the scheduled game start, never the padded market end date; other markets use their resolution deadline.
- PRICE — the recorded price is the live best ask a reader could have taken at emission. If the public order book is unavailable, the current outcome price is used and labelled; the wallet’s own fill price is retained separately.
- LEARN — every otherwise-qualified signal is scored against the Docket’s settled results that were already known at that instant. Market price is the prior. Wallet, market-family, wallet-family, price-band and conviction-size evidence adjust it. After warm-up, a correction below −3 percentage points is a PASS. Calls and passes are both recorded and resolved so the gate remains auditable.
- ONCE — while any decision on that wallet + market is still open, no second decision is emitted for it.
The results now change the next prediction.
The current model is derived from 708 settled observations across 634 markets. Market price remains the prior; demonstrated wallet and market-family performance adjusts it. Old calls never move, and every model version is retained.
Walk-forward check, not the live record: after the 50-result warm-up, the historical stream contained 641 calls at +$5.23 per $100. Applying each version of the learner only after its inputs had resolved retained 502 at +$9.02; the 139 signals it would have passed returned −$8.45. This is a time-ordered backtest, not a promise of future performance.
How the ROI figure is computed: $100 on every settled call at its versioned price recorded at emission (v2 uses the live best ask), resolution pays 1 or 0. Losses included. Gross of Polymarket's own fees — each call records the fee schedule its market declares.
That is the whole trigger. There is no discretion in it and no editor. The engine polls the public tape; every position that reaches the learning gate is written immediately to the decision record as CALL or PASS with its model id. CALLs enter the public ledger. Both kinds are resolved, so the learner is measured against what it rejected as well as what it selected. When a called market resolves, the same ledger row is marked hit or miss and its realised return is printed.
Calls are never backfilled. The record opened July 25, 2026 and cannot contain anything older than that, by construction.
Exhibit B — how a wallet gets rated
Right now: 100 wallets in the tracked cohort, 100 of them rated, refreshed aug 31, 2026 · 03:23 utc. The whole list is published, wallet by wallet, with each one’s own public settled record and the five components behind its score: the rated cohort →
A wallet is scored on its settled record only, on five components. A wallet without enough settled history is reported as UNRATED rather than given a flattering number, and an unrated wallet can never fire a call.
| Component | What it measures, and why |
|---|---|
| Edge | Realised ROI across settled positions only. An open position is an opinion; a settled one is a result. Unrealised gains count for nothing here. |
| Consistency | Win rate across settled positions, shrunk toward 50% when the sample is small, so a wallet that is 4-for-4 can never outrank one that is 180-for-300. |
| Breadth | How many distinct markets the wallet has actually traded. One lucky market cannot carry a score. |
| Discipline | A penalty for size dispersion. A flat-staking trader is demonstrating a process; a wallet with one 50× outlier is demonstrating a gamble. |
| Survival | Worst peak-to-trough drawdown on the realised curve — what the record actually felt like to hold, not just where it ended. |
The composite is the blend of those five. It exists to make one thing impossible: a wallet that dumped a large stake into a single correct coin-flip topping the list. That wallet tops every public leaderboard and following it is worse than useless.
Exhibit C — re-derive it yourself
Polymarket’s data is public and free. These are the same endpoints the engine uses. No account, no API key, no permission from us:
# the live tape — every trade, as it happens
GET https://data-api.polymarket.com/trades
# one wallet's complete fill history (money in and out)
GET https://data-api.polymarket.com/activity?user=0x…
# one wallet's positions, with resolution status
GET https://data-api.polymarket.com/positions?user=0x…
# the market: question, prices, resolution
GET https://gamma-api.polymarket.com/markets?slug=<market_slug>
GET https://clob.polymarket.com/book?token_id=<outcome_token>
# the current versioned learning model
GET https://oddsxray.com/data/learning-model.jsonTo check one call: take the call id and wallet from the ledger, pull that wallet’s activity, and confirm the position existed at the price and size we printed, at the timestamp we printed. Then pull the market and confirm how it resolved. Both halves are public; neither goes through us.
To check the whole record: do that for every line. That is the point of publishing all of them.
Two honest limits. Both endpoints above are Polymarket’s, so if they change shape or go down, our engine and your audit break together. And a wallet’s identity is an address, not a person — one human may run several, and we do not claim to know who anyone is.
The record is free. The calls, while they are still open, are $12 a week or $49 a month.
Everything that has resolved is on the ledger for nothing, permanently, no account. What you pay for is time: the call at the moment the rule fires, at the price it fired at, on your own live page.
See what it costs →