Field notes: the day the ranked list put the wrong site at the top

The query tile was a surface-to-air missile battery we'd had eyes on for months. Circular revetments, a berm pattern, an access road running in from the northeast at a particular angle. I knew that site the way you know a face. The task was simple to state and miserable to execute by hand: find every other site in the province that shared that signature.
So I ran a similarity pass across the AOI and got back a ranked list. Forty, fifty candidates, sorted by how close each chip scored to the query. I worked down from the top the way you're trained to, because rank order is supposed to mean something. The first result sat almost two kilometers from a provincial road, had the circular earthworks, had a road coming in from roughly the right bearing. On a screen at arm's length it looked like a sibling of the site I'd started from.
It was a quarry.
Not even a well-disguised one. The "revetments" were spoil rings from overburden removal, stacked in a circle because that's what you do with a conical pit and a front loader, not because anyone was protecting a launcher. The access road led to a weighbridge, not a staging area. I caught it because I went to pull a second date on the chip and the pit had grown six meters deeper in eighteen months, which a missile site does not do.
What stuck with me wasn't that the model got it wrong. Similarity retrieval doesn't know what a SAM battery is any more than it knows what a quarry is. It knows that two shapes occupy similar space in a feature embedding, and circular earthworks with radial symmetry and an access road are circular earthworks with radial symmetry and an access road whether a crew with artillery put them there or a crew with a Caterpillar did. The ranked list was doing exactly the job it was supposed to do. The job I'd stopped doing was reading the context around the match instead of letting the rank number do my thinking for me.
What it changed for me, practically, was the order I work a result list in now. I stopped treating position one as "the answer" and started treating the whole top band, say the first fifteen or twenty, as a set of hypotheses that all need the same five-second gut check: terrain context, a second date if the archive has one, anything in the surrounding tiles that would or wouldn't belong next to the thing I'm actually looking for. The list got me to candidates I'd never have found scrolling the archive by eye. It did not get me to a conclusion. That part's still mine.
A colleague who works coastal infrastructure told me a version of the same story with container cranes and grain elevators, two structures that apparently collapse into the same embedding neighborhood from a certain angle and resolution. I don't think that's a flaw worth fixing so much as a fact worth knowing going in. A tool built around marking one site and pulling back every lookalike across a wide search area, which is what led me to the pitch over at Broad Area Search, is explicit about this: it hands you candidates and leaves the verdict to you. That framing matches how the quarry day went. The list did the finding. I did the deciding, late, after I'd already half-believed the wrong answer.
The quarry's still there, still getting deeper. I check the date stamp on it now before I check anything else, mostly as a reminder to myself rather than because I expect it to have grown a launcher.
If your archive's too big to browse tile by tile and you've got one known site worth searching for, that's the exact gap Broad Area Search is built to close.