GeoLocator v3.0: Reasoning-First Geolocation
Archived — superseded by GeoLocator 4.0
This post is kept for the record. The accuracy, speed and model-size figures it originally quoted were never measured on a published benchmark, so they have been removed rather than restated. GeoLocator 4.0 is the current engine and the first version measured on our own published benchmark.
v2.0 was honest about a hard limit: a model that classifies a photograph into one of a fixed set of geographic clusters can never be more precise than the clusters themselves. To get closer than "roughly this part of the world", we had to stop treating geolocation as a classification problem.
v3.0 was the first version to put a reasoning step in front of the answer: the system described what it could see in the photograph — architecture, vegetation, road markings, signage — and only then committed to a location, so the evidence came out with the result.
From logits to an argument
Earlier versions produced a probability distribution over clusters and nothing else. v3.0 produced a written line of reasoning and a location, which made its mistakes legible for the first time: you could see which clue it had misread, instead of only that it was wrong.
Architecture: the reasoning pipeline
The classification head was replaced by a pipeline that chained visual feature extraction to a language-capable reasoning stage.
Before (v2.0)
Coordinates = the centre of the winning cluster
v3.0
Evidence first, then a location
Explanations you can audit
Because the model wrote down its reasoning before answering, that text could be shown to the user. In practice it read like a checklist — tropical planting, traffic on the left, a particular plate shape, colonial-era facades — and an investigator could accept or reject each observation on its own merits. That auditability, rather than any single accuracy number, is what made v3.0 useful, and it is the idea that survived into today's engine.
It also exposed the failure mode we spent the next year fixing: a fluent explanation can be completely wrong, and nothing in v3.0 told you which was which. GeoLocator 4.0 answers that with a separate decision layer and a confidence score that tracks correctness on our own published benchmark.
What came next
v3.0 was the engine behind the first GeoLocator dashboard, and it stayed in production until GeoLocator 4.0 replaced it in September 2026. Try the current engine.