We Watched 5,225 Crypto Bar Markets Settle Across Two Venues. One of Them Lies About Its Order Book.

2026-08-13 — Integrity Monitor — Data Post 1

Integrity Monitor — Data Post 1 By LowfieldLabs — dated kill decisions from the hypothesis ledger

For eleven days, 2026-08-02 through 2026-08-13, a read-only service of ours sampled the public order books of two venues — Polymarket’s CLOB and Kalshi — on paired crypto up/down bar markets: BTC, ETH, SOL, XRP, and DOGE, on 5-minute and 15-minute bars, every five seconds. It placed no orders. It just watched, recorded, and settled the books against each other.

This is the first publication from that monitor. Three findings, all of them bad news for a specific kind of trader, and all of them measurable by anyone with an afternoon and two API keys.

Finding 1: the venues disagree on who won, 3.9% of the time

Across the window we tracked 5,230 distinct paired bars — the same underlying, the same window, listed on both venues. 5,225 of them settled on both sides. In 205 of those, the two venues paid opposite sides: Kalshi’s UP winner was Polymarket’s DOWN winner, on economically identical contracts. That is 3.9% of settled bars.

Let that number sit for a moment. A cross-venue position constructed as a hedge — long UP on one venue, long DN on the other — is, roughly one time in twenty-five, not a hedge at all. It is a double loss or a double win, decided by two settlement feeds that do not agree about reality.

Our prior estimate, from smaller live-trading probes, was ~6%. We report both readings honestly: either the disagreement rate has regime-shifted downward, or the probes — small samples, selected windows — overcounted. Eleven days and 5,225 bars is a better sample than a handful of probes, but it is still one window in one regime. The monitor keeps running; that is the point of it.

Finding 2: one venue’s displayed book is largely fictional

We measure phantom liquidity: a displayed ask level of five shares or more that vanishes or requotes within 60 seconds, with no executions at the displayed price or a cent through it. A phantom is resolved against the public trade tape — if the tape shows the level traded, it was real; if the level disappears untouched, it was theater.

The averages across the window:

Nine in ten displayed ask levels on one venue evaporated without a trade. On the other, one in fifteen. That is a 14x difference in displayed-liquidity honesty between two venues listing the same economic contract. If your execution logic sizes against displayed depth — and almost everyone’s does — then on one of these venues your model of the book is mostly a model of a mirage.

Two caveats we apply to ourselves before anyone else does. Tape latency can misclassify very recent executions — a fill that printed in the last seconds of a level’s life may arrive after we scored it. And our derived Kalshi asks include ordinary quote improvement, which flatters Kalshi’s number slightly. Neither caveat moves a 14x gap into a 2x gap. The direction is not in doubt; the exact decimals might be.

Finding 3: the apparent margins are shrinking, and the remainder is increasingly fake

The monitor flags apparent cross-venue margins: moments when the two venues’ books imply a locked margin of 3 cents or more after our fee model (0.07·p·(1−p), taker-only — the conservative reading; if you can’t cross the spread, you don’t have a margin). Flagged instances per day:

The apparent arb capacity shrank roughly 3x inside eleven days. And the composition of what remains is worse than the count suggests: cross-reference Finding 2. A margin is two displayed prices, and if one leg’s price is phantom 91% of the time, the margin is mostly arithmetic on fiction. The opportunities that persist are increasingly the kind you cannot execute.

What this is and what it is not

This is not a free-money post. It is the opposite. The three findings together say: apparent cross-venue margins are mostly unexecutable, one venue’s displayed book is largely fictional, and the pair disagrees on who won one bar in twenty-five. Anyone running cross-venue arb on these markets without this data is not harvesting a spread. They are donating money, in a structured and repeatable way, to whoever is on the other side of the phantom.

We publish this because venue-integrity measurement is a public good and almost nobody does it. Venues report their own volumes. Traders report their own wins. Nobody with a stake in the answer measures whether the book was real or whether the two settlement feeds agreed. A research shop that publishes its kill decisions has no stake in the answer, so we built the meter.

The methodology, in one breath: read-only public APIs, no orders placed, ever. A phantom is resolved against the public trade tape, not assumed. Net margins use a taker-only fee model, 0.07·p·(1−p). The monitor’s known limits — tape-latency misclassification on very recent executions, ordinary quote improvement inside derived Kalshi asks — are stated above, because a meter that hides its own error bars is marketing, not measurement.

The kill ledger

LowfieldLabs keeps a kill ledger: one dated page per hypothesis, including — especially — the ones that died. The count stands at 1,232 hypotheses, 15 proven — we publish the kill decisions. The hypothesis this post kills is one we held ourselves: that paired bar markets across two major venues are close enough to identical that displayed cross-venue margins are executable. 205 opposite settlements and a 91% phantom share later, it is dead, and its page is dated 2026-08-13.

The monitor keeps running. Follow-up posts in this series will track whether the 3.9% disagreement rate holds or drifts, and whether the flagged-margin count stabilizes at its new level or keeps shrinking. If the numbers move, we will print the move.

This piece describes research methodology and public market data. It is not financial advice.