The biggest edge this machine has ever seen was 10.95%, locked in, four days to payout. This page is the story of why taking it would have been the mistake.
On August 24 the scanner priced a pair across two venues: YES on a Polymarket bitcoin market for $0.057, NO on Kalshi's "BTC above $80,499" for $0.60. Together, well under a dollar for a set that pays exactly $1, if the two contracts were really the same market.
The two contracts ask about different price thresholds, settled by different rules against different reference prices. If bitcoin finishes in the gap between the two strikes, both legs lose at once. That is not a hedge failing. That is two full-size directional losses wearing a hedge's clothes.
That 10 cents of "edge" is exactly the market's price for the gap. Nobody was leaving money on the table. The market was quoting a risk, and only a machine that reads settlement rules could tell the difference.
# matching.py, the tier doctrine
#
# Only "exact" is ever allowed to produce a tradeable kind.
# A big edge on a non-exact pair is not a bigger opportunity,
# it is a louder warning: the mismatch IS the edge, priced.For honesty's sake: this is one refusal out of 435 on the journal, picked because it is the largest. Most refusals are small and boring. The journal recording all of them, not the drama of this one, is the actual product.
The machine's newest lane pairs the same perpetual on two venues: short it where funding is rich, long it where funding is cheap, and collect the gap. The obvious way to pair them is by ticker. The obvious way is wrong. On one day in September, 21 of the 233 tickers the scanner follows traded at prices more than 3% apart across Hyperliquid, Bybit, and Binance. PURR on one venue was a different coin at $12 against $0.11 on another. PIXEL and OGN traded at four times their price on a venue where the market had been frozen after a delisting. A funding gap between two different assets, or between a live market and a frozen one, is not a carry. It is a number.
The rail is small and dumb on purpose: before a cross-venue pair can be sized, both order books must be read, and their mid prices must agree within 2%. A book that cannot be read cannot vouch, so the pair is refused rather than priced on a default. The same check now guards the spot carry: a spot market more than 2% from its perp is not that perp's hedge, whatever its ticker says. In the year of public funding history the machine was backtested on, the tickers this rail throws out were the biggest "winners" of the quiet months. They were never collectible. The rail took them out of the evidence before they could argue for real money.
Before Basis was allowed to trade anything, it was given a catalog of documented ways prediction-market arbitrage kills its operators. Every rail in the engine answers a line in that catalog.
Sybil wallets post fake mispricings, bait arbitrage bots, cancel, and force loss-dumps: $1.49M of attacker profit documented in published research. A fill is never truth until the first leg is mined, and implausible edges are treated as bait.
In published cross-venue pair audits, roughly a third of matched pairs settle by different rules. In February 2026 one market resolved opposite ways on two venues, both venues behaving exactly as written.
Measured arbitrage windows now average about 2.7 seconds, and 73% of captured set-arbitrage profit goes to sub-100ms infrastructure. Basis refuses the latency race and competes on reading rules, not racing fills.
Sources: arXiv 2606.16852, 2608.00666, 2604.24366, venue rule documentation, and on-chain records. The full catalog is applied as a review checklist on every build.
Set up a cross-venue pair and watch the engine decide. Defaults recreate the August 24 refusal.
Running a bot that fills fine but still leaks money? The audit runs your pairs and fills through this same catalog and names every trap, with the cost of each. If it finds no concrete leak, you pay nothing.