About & limitations

What PolarTrap claims

PolarTrap detects unusual patterns of coordinated-looking antagonistic interaction and shows the evidence for human review. It never states that accounts are run by the same person, and its labels are signals, not proof.

How evidence is counted

• All signals derived from the interaction itself (recurring arguments, reply latency, alternation, simultaneity, reply density) count together as one key, because two real people arguing actively can produce all of them at once.
• A second, independent key must come from outside the interaction: writing style with topic words removed (and long-term posting schedule where available).
• Thresholds are percentiles among comparable real antagonistic pairs in a frozen reference corpus, fixed before we looked at any test results. Reference pairs are a distribution for comparison, not confirmed innocent accounts.
• A percentile is not a probability of coordination.

Known limitations

• Domain shift. Our test pairs were written live in a sandbox during one event, while the reference corpus comes from Bluesky. False-flag rates measured in the sandbox apply to the sandbox setting and may not carry over to Bluesky. Sandbox personas also write far less than typical Bluesky accounts, so their writing-style comparison often uses widened reference matching (shown on every evidence card).
• Staged vs. intense arguments. If two different people stage an argument and leave only interaction traces, PolarTrap may be unable to tell it apart from an intense genuine argument. Interaction pattern alone is never enough for a flag.
• Low-text accounts. Writing style needs at least 500 words per account (1,500 for full confidence). Below that, the answer is Insufficient evidence, not "normal".
• English only in this version.
• Sophisticated actors who deliberately vary their writing style can evade the style key.
• Small pilot. Detection rates come from a small exploratory pilot and are reported with uncertainty intervals, not as validated performance.

Ethics

• Read-only on Bluesky: we never post, reply, like, follow or create accounts there.
• Accounts are shown as pseudonyms (acct_ + keyed hash). Real handles are visible only to the team in a password-protected view.
• Sandbox volunteers gave consent, wrote all text by hand, and are stored only under pseudonymous IDs.