Geolocation and chronolocation
Establishing where a photograph or video was taken and when, from the content of the image itself. The two are one discipline because they trade against each other: fix the place and the sky tells you the time; fix the time and the sky tells you which places are impossible.
This is the craft skill at the centre of open-source investigation, and the spoke covered the field for six weeks without holding it. ironsight plots coordinates that feeds supply and verifies none of them, so the corpus had the aggregation half and none of the verification half.
How it is actually done
The method in sun-shadow-geolocation is representative: compute the sun’s position for a candidate date, time and place; read the shadows and the sun’s bearing in the frame to fix camera orientation; use that orientation to eliminate regions rather than to search them; then settle the remainder on distinctive built detail: a lane count, a row of palm trees, signage. Landmarks, terrain, satellite comparison and street-level imagery are the other standard inputs.
Elimination is the load-bearing move. A constraint that rules out everything south of a river is worth more than a resemblance that suggests one street.
Where it sits in a formal investigation
glan-bellingcat-methodology places geolocation and chronolocation in Phase VII, Verification & Analysis, alongside corroboration and content description, feeding the Incident Assessment Report where “this verification process must be clearly displayed.” The methodology assumes the skill and regulates its documentation — the reasoning has to survive a reader, and potentially a court.
That documentation demand is the same one berkeley-protocol states as ¶25 explainability. A geolocation is not finished when the analyst is convinced; it is finished when someone else can follow the steps.
The automation question
llm-geolocation-test is the measurement: 20 models over 25 unpublished photographs, scored against Google Lens, with three OpenAI reasoning models beating that baseline and every model hallucinating at some point. Models are useful for noticing (multilingual signage, small cues in urban scenes) and unreliable for concluding.
The structural mismatch is worth naming. The craft’s value comes from constraints that force an answer and can be checked; a model returns a place without the constraint chain that produced it. That is the opposite shape from what both standards above require, which is why “the model said so” cannot close a geolocation the way a shadow angle can.
Related
sun-shadow-geolocation · llm-geolocation-test · glan-bellingcat-methodology · berkeley-protocol · bellingcat · conflict-monitoring · ironsight · osint