Logical-implication arbitrage, with an LLM finding the logic
If market A happening logically requires market B, then A can never price above B — any violation is riskless profit. An LLM found over a thousand implication pairs. The books never once mispriced them.
The hypothesisMarkets are priced one at a time, so cross-market logic must occasionally break: "X wins the championship" trading above "X reaches the final," thresholds out of order, subsets above supersets. Regex catches numeric ladders; a language model can catch semantic implication at scale, opening a whole class of riskless trades.
The test
Every few hours, active markets were clustered by topic and an LLM named pairs where YES on one logically necessitates YES on the other (necessity only — correlation explicitly excluded). Each proposed pair was then judged at the live order books: a real violation means you can sell the stronger claim's bid and buy the implied claim's ask for a locked profit. Every pair was also validated at resolution, scoring the LLM's logic itself.
The result
The model's logic was good: of 876 resolved pairs, 849 implications held (~97% — the failures were mostly subtle resolution-criteria mismatches, not reasoning errors). The market's pricing was better: across 1,065 pairs, the number with an executable violation at the books was zero. Along the way the probe also documented the trap that makes this class dangerous: mutually-exclusive bucket markets (temperature ranges, exact scores) look like implication ladders to a language model, and its most exciting "violations" there were anti-arbitrages that lose money with certainty.
Conclusion
The market does not misprice logic — at least not in sizes and durations visible to an every-few-hours scanner. Consistency across related markets is evidently maintained by the same professional flows that keep Up+Down above $1. The experiment retired the LLM-arbitrage thesis and reinforced the doctrine those flows keep teaching us: structural free money does not survive in patrolled books, and an AI that can find real logic still can't find anyone mispricing it.