dutchbook — Prediction Market Coherence Monitor 2026

A prediction market price is a probability, and probabilities have to obey arithmetic. "Bitcoin above $150k" cannot happen without "Bitcoin above $120k" also happening, so the first can never be priced higher than the second. Eight candidates of whom exactly one must win have to sum to a dollar. When prices break rules like these, the contradiction is a fact about the venue that needs no opinion about the world — and a set of bets exists that wins money whatever happens. That set is called a dutch book.

dutchbook extracts 67,617 such relations across 264,128 Polymarket markets — 85.4% of the active catalog — checks them continuously against a live order-book feed, and records how long each violation survives. The headline finding is a median lifetime of 15 seconds across 2,248 closed episodes: half of every logical inconsistency found was gone before you could finish reading about it, which is far too fast to be a person and says the correction is either automated or was never a correction at all.

The more useful half of the project is what it refuses to claim. Only 3.9% of episodes survived being priced against real depth and fees, and the 90 that did trace back to just two relations — both of which turned out to be bugs in this system's own extractor. The executable-arbitrage count in the report is therefore published as zero.

• Catalog: 864,077 markets ingested, 14,951 active — 85.4% of the active catalog appears in at least one constraint.
• Constraint graph: 67,617 relation edges and 24,201 partition groups — threshold ladders 65,249, nested deadlines 2,238, human-reviewed LLM proposals 128, explicit complements 2.
• Lifetime: median 15s, p75 27s, p95 5.5m, max 3.6h over 2,248 closed episodes.
• Apparent vs confirmed: 2,308 episodes, 90 survived live depth and fees — 3.9%. No partition violation was ever executable, because an n-member basket needs n simultaneous fills and pays the spread n times.
• Detection latency: median 0.06s — the screen is event-driven off the CLOB WebSocket, not polled; a book update marks only the constraints that market participates in.
• Dashboard: 6.7s cold, 90 ms warm after replacing a TTL cache that was pure overhead with stale-while-revalidate.

Every figure in the report is computed from the database at run time by pnpm report; none is written by hand, so a stale number cannot survive a regeneration.

Median net edge on the confirmed violations was 30.66¢ per share, maximum 59¢. A risk-free thirty cents on a dollar-denominated contract, sitting untouched for sixteen seconds, is not something a venue with real participants leaves lying around — so the report said so before checking, and then checked.

The 90 confirmed episodes were 2 distinct constraints, each re-detected 45 times as it flickered across the threshold. Both were wrong. "OpenAI does not IPO by 2026" was recorded as entailing "does not IPO by 2027" — true for a positive event, backwards for a negated one. The second didn't even need the subtlety: a deadline of Dec 31 recorded as entailing one of Sept 30 is wrong on its face.

This was the second appearance of that class of bug. The first put an inverted direction on 888 markets — the ladder extractor read "hit 35%" as an upward threshold when Polymarket's own criteria resolve Yes at or below the value — and produced a $435 "risk-free arbitrage" that paid zero in one of three states. Nothing alerted either time, because nothing was broken: every component downstream did its job flawlessly on a false premise. A reversed entailment is the worst failure mode this system has precisely because it does not fail loudly — it produces a confident, well-formed, fully-priced answer.

That is also why LLM-proposed edges require a recorded human verdict before entering the graph. A missing edge costs an opportunity nobody notices; a reversed one costs money with total confidence.

• The window is hours, not a month: 4.7 hours across two calendar days. Distributions have n=2,248 behind them and are worth reading; every rate, weekday effect and intraday pattern is unavailable and is reported as unavailable rather than estimated.
• The size/lifetime hypothesis is unanswerable here: Spearman rho is +0.157 — larger contradictions lasted longer, and the sign holds inside each constraint type, so it is not Simpson's paradox. But 90.3% of episodes lasted under two 60-second intervals, so for nine in ten the recorded lifetime is the sampling rate rather than the market.
• 15 seconds cannot distinguish arbitrage from jitter: either automated traders close these gaps faster than a human can act, or independently drifting quotes crossed a threshold and crossed back. The measurement cannot tell those apart, and the report picks neither.
• The fee model has three named weaknesses: the taker rate varies 0.04–0.07 by a category no endpoint publishes, so the highest rate is applied to everything — deliberately under-confirming. The cost register is printed from source rather than restated, so an assumption cannot change without the table changing with it.
• Coverage is of markets, not of logic: cross-event implications are almost entirely absent, and conditional or compound structure has no representation in the constraint language at all. Every "no violation" means "none among the constraints we know about".