PolarTrap

Bot detectors find accounts that agree too much. PolarTrap looks for accounts that disagree suspiciously well, and tells you when it can't tell.

How it works

1. We import the public reply tree and each participant's recent public posts (read-only).
2. A language model labels each reply as agree / disagree / neutral. Failures are counted as unknown, never guessed.
3. For pairs that argue, we compare two independent kinds of evidence with a frozen reference corpus of real Bluesky arguments: interaction pattern (one key, however many signals feed it) and writing style with topic words removed.
4. A pair is surfaced for human review only when the evidence is extreme against that reference. When data is missing, the answer is Insufficient evidence.

Accounts are shown as pseudonyms. A percentile is a comparison with the reference corpus, not a probability of coordination. PolarTrap never claims that accounts are run by the same person. See measured results.