We’ve written before that the map is made of open tabs — that Thyseus runs object detection inside contributors’ browsers, on public camera feeds nobody was watching, and never ships the video anywhere. That note ended on a phrase it didn’t cash out: the software’s job is to let all the tabs agree on what they were looking at. This is the note about the agreeing. It’s the harder half, and it’s the half that decides whether you have a map or just a pile of guesses.
Start with the uncomfortable question, because a skeptical reader gets there in about four seconds: why would you trust a map that strangers’ laptops drew? You don’t know their hardware. You don’t know their network. You don’t know if the tab watching a bridge at 3 a.m. is running a good detector or a potato. And you’d be right to ask — because any single detection coming off any single browser is not a fact. It’s a rumor.
One tab is a rumor
A detection model running in a browser is small on purpose. It has to boot in a tab, share a GPU with everything else, and keep up with a live frame rate, so it is nobody’s idea of a research model. Point it at a real public feed — compressed, rain on the lens, a low sun blowing out half the frame, a JPEG artifact that looks exactly like a sedan — and it will make mistakes. It will call a box truck a bus. It will see a car in a shadow. It will miss the motorcycle entirely.
None of that is a bug to be stamped out, because you can’t stamp it out; it’s the standing condition of cheap perception at the edge. A better single observer is not the move. Even a perfect model on a perfect feed only ever sees one vantage, and the moment you’re stitching thousands of feeds together, the interesting failures aren’t blurry pixels — they’re confident, plausible, and wrong. A rumor delivered crisply is still a rumor.
So the design question is not “how do we make each tab right.” It’s “how do we build something trustworthy out of observers we’ve already conceded are individually unreliable.” That’s not a computer-vision problem. It’s an old journalism problem, and it has an old answer.
Corroboration, computed
A newsroom doesn’t trust a tip because the tipster sounds sure. It trusts a story when independent sources, who don’t know each other, describe the same thing. Thyseus works the same way, mechanically. A detection isn’t a claim about a pixel; once you fold in the camera’s known position and field of view, it becomes a claim about a place — this vehicle, at roughly this latitude and longitude, at this second. And a claim about a place is something other observers can independently confirm or contradict.
So the pipeline stops treating detections as answers and starts treating them as votes. Several tabs watching overlapping feeds — or the same corner from different cameras — each compute where they think an object is. Cluster those computed positions. Where a real car sits, the estimates pile up in a tight knot: different machines, different models, different angles, landing in the same few meters. The lone tab that saw a car in a shadow produces a point with no neighbors, and a point with no neighbors doesn’t make it onto the map as a car. It makes it onto the floor.
Confidence isn’t the number the model prints next to a box. It’s what’s left after the observations that don’t agree with each other have quietly fallen away.
That reframing is the whole trick. The confidence Thyseus shows you is not a detector’s self-reported certainty — those are notoriously overconfident, and a model that’s wrong is often most sure. It’s an emergent, cross-checked quantity: how many independent observers, with uncorrelated failure modes, converged on the same place. A pin backed by seven tabs from four cameras is a different kind of object than a pin backed by one, and the map is honest about which is which.
Why independence is the load-bearing word
The reason this works — the reason it isn’t just averaging noise — is that the observers are independent. Different contributors, different GPUs, different builds of the model, different physical angles on the scene. When errors are independent, they don’t line up; two machines rarely hallucinate the same car in the same spot at the same instant, so agreement actually means something. Averaging correlated errors just launders them into a confident average that’s wrong. Averaging independent ones cancels them out. The diversity of the crowd isn’t a tolerable cost of doing it cheaply — it’s the exact property that makes the cheap way trustworthy.
Which also tells you where the real threats live, and they’re not blurry frames. They’re correlated failures: a single contributor spinning up a hundred tabs to stuff the ballot, or a bad actor feeding doctored frames to manufacture a pin. Fifty copies of the same lie is not consensus — it’s one source wearing fifty coats. So agreement has to be weighted by genuine independence, not raw count: the leaderboard and the reputation attached to a contributor aren’t gamification bolted on the side, they’re part of how the map decides whose corroboration is worth what. The engineering that matters here is less “detect the car” and more “make sure the fifty witnesses are actually fifty people.”
Being honest about a thin crowd
The same discipline that keeps code-results from dressing a correlation up as a proof applies straight to the map. Consensus needs a crowd, and the crowd is uneven. A busy interchange watched by a dozen tabs produces pins you can lean on. A rural camera with one sleepy observer produces a single unconfirmed estimate — and the right thing to do with an unconfirmed estimate is to show it as unconfirmed, faint and provisional, not to promote it to fact because it was the only thing we had. A corner nobody is watching is simply dark. We’d rather show you the seams than paint over a thin spot with false confidence, because the moment a map starts inventing certainty it doesn’t have, you can’t trust the parts that were real either.
That’s the quiet reason a two-person studio can ship something that looks like infrastructure. We didn’t out-engineer the big cameras-and-GPUs approach on its own terms. We changed what “trust” is made of — moved it off the quality of any one sensor and onto the agreement between many cheap ones — and agreement is a thing you can buy with attention instead of hardware. The map is made of open tabs. It’s held together by the tabs agreeing, and it’s honest, out loud, about every place they haven’t agreed yet.