AI does the searching.
Statistics does the deciding.
Beating a market this efficient is a search problem before it is a modelling problem. The ideas worth testing are rare, and they are buried in an enormous space of plausible ones that turn out to be noise. We run that search with AI agents, continuously — and then make it very hard for anything they find to reach a live position.
The space of things that might predict a game is effectively infinite and almost all of it is noise. Agents scout new sources, read the graded book for conditions worth suspecting, and draft testable hypotheses continuously — at a volume no analyst could sustain, and without the ego that makes a person defend last week's idea.
One hard constraint makes the rest trustworthy: a deterministic engine computes every projection, probability and price, and no language model may compute or alter a number. Agents decide what to investigate and how to describe it. They never do the math.
Nothing is promoted because it sounds compelling. It is registered, tested walk-forward, corrected for multiple testing, and kept out of the book until a forward record earns its place.
From feed to verdict.
6 upstream providers wired in · 30 data points live and measurable · 430 points in the index · 117 approved and queued to build
- 01ACQUIRE
Every feed we can source is polled continuously and scored for whether it could plausibly move a price.
- 02REGISTER
A hypothesis is written down — claim, target sample, expected direction, exact filter — before the data exists.
- 03TEST
Walk-forward against real closing prices, corrected for the fact that hundreds of tests are running at once, alongside null controls.
- 04SHADOW
Anything that survives runs forward with no capital behind it, so we find out whether it holds outside the sample that produced it.
- 05GRADUATE
A forward record clearing ROI ≥ 6% on n ≥ 40 is the only way anything reaches the book.
- 06RE-GRADE
Every live rule is re-tested nightly. When the forward record stops paying, the rule dies — currently 6 of 13 are flagged for removal.
This one is designed to find nothing.
Every program is committed as a record like the one beside this — the claim, the target sample, the direction we expect, and the exact filter that selects the fills — before the data exists. It cannot be edited afterwards without leaving a dated entry in the log.
We run null controls on purpose. A coinflip market should show no edge. If this test ever comes back positive, it is not a discovery — it means our harness is broken and everything it has ever produced is suspect.
SEE ALL 40 OPEN PROGRAMS →{
"id": "xs-control-fair",
"sport": "cross",
"title": "NULL CONTROL — liquid coinflips are fairly priced",
"hypothesis": "45-55¢ liquid markets should show NO edge. If this fires, the harness is wrong.",
"min_n": 150,
"expect": "above",
"test": {
"type": "corpus_bucket",
"when": {
"price_min": 45,
"price_max": 55,
"volume_min": 200
},
"side": "yes"
},
"added": "2026-08-10",
"status": "queued"
}Named inputs, not a logo wall.
A research lab is only as good as what it ingests, and a list of hostnames is something you can go and check. These are the feeds currently wired into the model.
live game state, rosters, box scores
Statcast — pitch velocity, spin, contact quality
historical play-by-play back to the 1900s
NBA/CBB/CFB/NFL results and archived closing lines
settled event markets — our price reference
park-level wind, temperature, humidity
Hunt the edge: advance the proven signals to real plus-money execution, mine one new watch, and keep the company/brain/product compounding.
Make the company, the business, and the model better — every day. Find an edge.
Rewritten each morning from the state of the book, the open programs and last night's results.