Shared narratives
Model-specific cosine matching can find nearby meanings after calibration. A separate shingle-overlap channel identifies near-copy text.
Many voices,
one script.
A closer read of the Russia / Ukraine information space.
Chorus looks for accounts repeating the same narrative, close in time, again and again. Looks coordinated? Inspect the evidence before drawing a conclusion.
Read a post in context or inspect the live evidence. The notebook below keeps our labelled demonstrations separate from current observations.
Local CPU embeddings need model-specific calibration. Missing model files use disclosed lexical overlap; the synthetic demo is not measured accuracy.
Start an investigation
No uploads saved to disk. Text may remain in a bounded in-memory cache. Embeddings run locally by default; an explicitly configured HTTP backend sends text to that service.
Evidence notebook
Pick a sample or bring your own posts. Your results will appear here.
A single post tells you little.
Repeated co-activity tells a different story.
What counts as evidence?
Repeated synchrony
The provisional default requires at least two independent pairs within 15 minutes per account link. A single viral post is not enough. Evaluated ablations make removed safeguards explicit.
Connected behaviour
At least three accounts, with evidence you can inspect. Account similarity is shown for context, but contributes no score unless the evaluated configuration retains it.