Evidence
Every hour of every night in 2026, graded against the identity. Figures in basis points; lower is tighter.
When the overnight book started being quoted
An identity can only be checked when both legs are actually quoted. Before 2025 the dark window is empty, which is why the study starts where it does — and why nobody has audited this yet.
How often each identity breaks
Session is the same arithmetic measured inside US hours, where arbitrage enforces it. Treat it as a floor on measurement noise rather than a clean control: hourly sampling across two legs in a fast market manufactures some residual of its own. The headline here is the absolute breach rate, not the ratio.
| pair | required | nights | session | dark | >20 bps | p99 | worst |
|---|
The trade we did not find
If a breach closed by morning it would be an arbitrage. We entered every reading past 40 bps and held to the open. Capture is what the breach gave back; cost is two sides of taker fee on gross notional — four dollars working per dollar of signal on a 3× pair, two on a mirror.
| pair | +1h | +3h | to the open | n | costs | net |
|---|
Every net column is negative. The breach is real, it is measurable, and it is not a trade — which is exactly why it belongs in a pre-trade check rather than a strategy. You do not harvest these. You avoid paying them.
It is not stale prints
The obvious objection: a wide reading is just an hour where one leg did not trade. Median turnover inside the leg's own bar, split by how wide the reading was. Wide bars are as busy as quiet ones, or busier.
| pair | quiet bars (<20 bps) | wide bars (≥60 bps) |
|---|