Both sites, or it didn't happen
Ten milliseconds is the whole argument
The maximum possible separation sets the coincidence window, and everything outside it is discarded by definition.

Geography as physics
LIGO has two detectors: one outside Livingston, Louisiana, and one near Hanford, Washington. The straight-line distance between them is roughly 3,000 kilometres. Light — and gravitational waves, which travel at the same speed — takes about ten milliseconds to cover that distance. That single physical fact is the foundation of every detection claim the collaboration has ever published.
The argument works like this. A real gravitational wave arrives as a planar wavefront crossing the Earth. It will reach one detector before the other, or both simultaneously if it arrives from exactly the right direction, but it cannot arrive at the second site more than ten milliseconds after the first. That ceiling is set by the separation and by the speed of light — neither of which is negotiable. So any two events that fall within that window are candidates; anything outside it is not a coincidence at all, by definition, no matter how large the signal.
The maximum possible separation sets the coincidence window, and everything outside it is discarded by definition.
This is not a filtering heuristic or a statistical convenience. It is geometry.

What the window actually discards
The coincidence window cuts in both directions. Tighten it and you reject real signals that arrive close to the worst-case light-travel time; loosen it and you let in noise coincidences that happen to overlap by chance. Ten milliseconds is the maximum physically permitted, so it is also the maximum scientifically defensible. LIGO's published analyses typically apply a window of ±10 ms around any trigger at one site, then search the data from the second site for a matching event inside that range.
The background is estimated by time-sliding one detector's data stream past the other's by intervals far larger than ten milliseconds — hundreds of seconds, stepped repeatedly — so that no real coincident signal can contaminate the background sample. Whatever survives that process is treated as accidental coincidence caused by noise alone. The rate of such accidental pairs, measured across thousands of time slides, becomes the false-alarm denominator against which any real candidate is judged.
Key numbers
The first confirmed detection, GW150914, arrived at Livingston 6.9 milliseconds before Hanford — well within the window, and consistent with a source direction that has since been localised on the sky by the LIGO and Virgo collaborations. That 6.9-millisecond gap is not a footnote; it is part of the evidence. A noise transient at one site has no reason to produce a matching transient at the other site 6.9 milliseconds later. The probability of that happening by accident, given the measured noise backgrounds, is what the false-alarm rate quantifies.

The single-detector problem
A single interferometer, however sensitive, cannot make this argument at all. Local disturbances — seismic events, acoustic coupling, instrumental artefacts — produce transients that are indistinguishable from a gravitational wave signal by shape alone. The coincidence requirement is precisely what separates "a disturbance in one vacuum chamber in Louisiana" from "a spacetime disturbance that crossed the Earth." Without a second site and a physically motivated time window, the measurement has no way to rule out the former.
This is why LIGO's sensitivity curve matters in tandem with coincidence rather than instead of it: sensitivity tells you what signal amplitudes the instrument could detect; coincidence tells you whether what it detected is real. The two requirements are not alternatives. A detector can be arbitrarily sensitive and still produce no credible detections if it operates alone, which is why the two-site architecture was built into the project from its earliest design documents.
Ten milliseconds is a small number. But it is the entire logical structure of the claim that gravitational waves have been observed — not inferred from a single channel, but confirmed across a continent, inside a window that physics itself drew.