The queue was the arbitrage. The gap was just where it showed.

Both venues publish their complete trade tapes. On the coarse grid a pegged pair trades at a handful of prices a day, all multiples of 0.0001; the day the tick narrows, dozens of five-decimal prices appear. That signature dates the first venue's change to 14 April 2026 and the second's to 8 September, timed at 07:00 UTC from our order-book recording. The sample therefore has four regimes: both coarse until 13 April, asymmetric for 147 days, the change day, and both fine from 9 September. The first and last are controls for the middle.

We also ran the same count backwards over each venue's entire published history, one from December 2018 and the other from September 2021. Every one of the 143 venue-months before the changes returns 100.0% of trades on the 0.0001 grid. Each venue changed this tick exactly once in its published life, and the coarse setting survived every stress episode in between, including March 2023 when the asset traded down to 0.87. The rent we measure below was not a transient oversight; it was the standing configuration of this market for five to seven years.

Trade tapes give executed prices, which bounce between bid and ask, so a raw print-to-print comparison is dominated by bid-ask bounce. We compare one-second volume-weighted average prices instead: for each UTC second in which both venues traded, each venue's VWAP is computed and the gap is their difference in basis points. This keeps the measurement free of any order-book data, so it runs across the whole window.

Month by month the asymmetric share sits between 40% and 47%. There is no single episode driving it. A trader watching these two venues would have seen, on four days out of ten by time, one venue's price at least half a basis point from the other's, on a pair whose fundamental value cannot move.

The direct way to take a cross-venue gap is to buy on the cheap venue and sell on the dear one, crossing the spread on both. Without book data this cannot be simulated exactly, but it can be bounded from above using only prices that traded: the evidence that an ask existed at P is an aggressor-buy print at P, and the evidence that a bid existed at Q is an aggressor-sell print at Q. For each matching interval a round trip is possible whenever one venue's bid exceeds the other's ask, sized by the printed volume on each leg and a $10,000 clip, with $20,000 of capital since inventory must sit on both venues.

Two things follow. First, the opportunity is an artefact of the matching window: inside one second the price itself moves, so a bid on one venue above an ask on the other is usually the price having drifted between the two prints rather than two prices existing at once. Second, and decisively, the bound is higher after the window closed at every tolerance — 0.5% a year against 0.1% at ten milliseconds. Whatever residue survives strict simultaneity is created by trade frequency, which the finer grid increased, not by the tick asymmetry. The coarse grid contributed nothing a market order could take.

The other way to take the gap is to be passive on the coarse venue: rest a quote at its touch and, when filled, hedge immediately on the fine-grid venue. This needs resting depth, which the coarse venue happens to publish as a daily 400-level order-book file retained for about three and a half months. We retrieved 91 coarse-grid days, re-priced the quote whenever the touch moved, and hedged each partial fill 30 milliseconds later. The result turns on one modelling choice: where in the queue the participant stands each time the order is re-priced.

The two rows are different businesses, not two sizes of the same one. The newcomer's row is a strategy: about 5.7% a year on the $20,000 the two legs require, before competition and before any cost of moving coins between venues. The incumbent's row is a bound, not a strategy: filling $14.6 million a day means absorbing 22% of everything the venue traded, which requires winning the race to the front of every price level against $13 million of standing orders. What the row establishes is where the money was, not that it was available.

Sections 2 to 4 leave a puzzle: the gap was there 41.5% of the time for five months and nobody could take it. The resolution is that the coarse grid's extra spread was never a cross-venue mispricing. It was a rent paid by that venue's own takers to the makers holding its queue, and it can be measured directly by splitting each trade's effective spread into the part the maker still held five seconds later (realised spread) and the part the price moved against them (price impact, the compensation for informed flow).

Because both venues quote a one-tick spread almost always, the mid can be reconstructed from the trade tape alone — an aggressor-buy print at P means the ask was P and the mid P minus half a tick — which lets the decomposition run over every trade on both venues for the whole window. We verified the identity against the coarse venue's own published order book on 86 days: the two methods differ by 0.005 bp on the realised spread.

This table contains a replication. The two venues cut their ticks five months apart on the same asset and produce the same result: on a coarse grid the maker keeps about two thirds of a one-basis-point spread and the adverse-selection share is about a third; on a fine grid the maker keeps a few hundredths of a basis point and the adverse-selection share is about two thirds. Each venue's realised spread falls by a factor of about twenty at its own change, while price impact barely moves. The extra 0.9 basis point on the coarse venue bought its makers no extra protection against informed flow. It was a transfer from that venue's takers to the front of its queue, imposed by the grid.

That reconciles the gap with the money. To collect the rent one must hold queue priority in a $13 million line, which is a capital and latency position, not a trade. A newcomer at the back is filled disproportionately on the fills nobody wants. And a cross-venue arbitrageur who hedges every fill immediately hands most of the rent straight back to the other venue: even from the front of the queue the hedged edge was 0.40 bp against a realised spread of 0.62 to 0.66 bp. The rent was payable in full only to a single-venue market maker willing to carry inventory. There was money in the coarse grid, it was substantial, and it was not in the cross-venue gap that the coarse grid made visible.

Against a three-month baseline from the venue's own tape, the tick change took its quoted spread from 1.00 to 0.12 bp, resting depth at the touch from $13.5 million to $2.8 million, daily volume from $70 million to $300 million, and trades per day from 51,400 to 136,500. The larger venue's volume rose about a fifth over the same interval, so roughly a twentieth of the fourfold rise is market-wide.

The last number matters most for price discovery. A book whose spread is one tick and whose tick is a full basis point has a mid that cannot move by less than half a basis point, and a stablecoin's fair value almost never moves that much. On the coarse grid the venue's mid was not a price; it was a rounding of the other venue's price, which is why the cross-venue gap existed at all and why it carried no information. Only once both venues were on the fine grid does a vector error-correction model return stable information shares — the larger venue leading at every sampling interval from one second down to 100 milliseconds — and only then does the ordinary machinery of lead-lag measurement have anything to measure.

• One event per venue. Neither venue has changed this tick at any other point in its published history, so the window cannot be replicated on another date. What generalises is the mechanism, not the magnitudes.
• Resting depth covers 91 of the 147 days. The coarse venue's published order book reaches back about three and a half months; the first eight weeks of the window have no depth data from any free source. The mechanism is measured over the full 147 days from tapes alone.
• The hedge is modelled. It is insensitive to delay up to a second and loses 4% when displayed depth is enforced, but no replay can model other participants hedging the same flow at the same instant. The $3.14 a day is an upper bound for that reason.
• Inventory, not execution, is the real constraint. Every capture figure assumes inventory already sits on both venues; rebalancing means an on-chain transfer that can be suspended in exactly the dislocations that would make a window valuable.