Six ways of acting together
The method looks at the authors in one topic’s window. Two accounts that take the same action inside the window form a pair for that way of acting, and the pair’s weight is the number of times, discounted by how widely the shared thing was being shared in the window.
The windows are not taken on trust. In each run the method also tries windows of 10 seconds, 30 seconds, one minute, five minutes and fifteen minutes, and the result is written into the finding.
| Action | Window, by default | What it suggests alone |
|---|---|---|
| The same repost | 60 seconds | Little. Fans of an account repost together within seconds too, and a pair is not a finding. |
| The same text | 5 minutes | A template, when the words match, or a rewording close enough in meaning, in the same language or across languages, because campaigns reword their messages. Captions reused on short videos count. |
| The same link | 60 seconds | Little when the link is popular: a link that very many accounts are sharing counts for less, so one viral link cannot make a network. |
| The same hashtag | 30 seconds | The same: a hashtag the whole topic is using counts for less. |
| The same hashtags in the same order | 5 minutes | A sequence of three or more hashtags, by default, in the same order. It is the trace that holds on video platforms, where there is little text to compare. |
| The same media | 5 minutes | The same image, video or audio, matched by its fingerprint, posted by both. |
The same second
Coordination runs every five minutes for each trend, on the previous topic snapshot.

01
Pairs in each way
For each of the six ways, every pair of accounts that acted inside the window is counted, with its weight.
02
Chance removed
Pairs of weight one are dropped. A pair is kept only when its weight is beyond what a noise-corrected model of ordinary activity predicts, by more than 2.32 standard deviations by default, about one chance in a hundred. The fallback filter keeps the strongest twentieth of each way’s pairs.
03
Isolates dropped
Accounts left with no pair leave the picture. Which filter ran, and its settings, are recorded with the finding.
Eleven measures
Each group found is scored on eleven measures, each from 0 to 1. A measure whose inputs are missing is left empty, its weight is shared among the others, and the report says so. A measure counts as an indicator when it passes its own mark, 0.5 by default. The weights below are the defaults and sum to one.
01
Synchronised bursts · 0.10
The share of the group’s posts that fall in bursts of 20 or more within 60 seconds.
02
Text templates · 0.15
The share in clusters of five or more matching texts.
03
Aligned narrative · 0.15
The share of posts whose nearest match is a rewording close in meaning, not a copy.
04
Repeat offenders · 0.15
Members found in groups in two or more earlier topics within 90 days.
05
Cross-platform spread · 0.10
The same cluster carried on two or more platforms within the window, by members or accounts linked to them.
06
Co-repost density · 0.10
Of all the possible pairs among the members, the share joined by a kept pair for reposting the same post.
07
Inter-arrival gaps · 0.10
The share of member pairs whose typical gap between shared posts is shorter than the topic’s ordinary gap. It finds schedules of minutes, not only bursts of seconds.
08
Schedule regularity · 0.05
How strongly the group’s combined posting repeats at a fixed interval over 24 hours, read in five-minute steps. It is a property of the group, not of each member’s own schedule.
09
Creation cluster · 0.05
Members’ accounts created close together. It needs profiles, and is left empty when fewer than one author in five has one.
10
New or small accounts · 0.03
The share of members with fewer than 100 followers or an account younger than 30 days, among the members that have a profile. It is left empty when fewer than one member in five has one.
11
Repost-only accounts · 0.02
The share of members with at least 10 posts in the window and none of them original.
From pairs to a group
Pairs become a named group in three steps, and the topic’s score and label follow from the groups.

1
Pairs
Every pair that acted together inside the window, in each of the six ways.
2
Chance removed
Only pairs beyond what ordinary activity explains remain; accounts with no pair drop out.
3
A group
Communities are found in each way and across all six together. A community is a group at five accounts or more and at least one in twenty of the topic’s posts. A text group with five or more different rewordings is a campaign of rewordings, with its languages. If no community qualifies, the whole topic is scored as one group by the same method.
04
The score
Each group’s strength is the weighted mean of its measures; the topic’s score adds each group’s strength times its share of the posts, from 0 to 1.
05
The label
One of four, against two lines, 0.70 and 0.30 by default. At or above 0.70: likely coordinated with at least three indicators, otherwise a lead. At or below 0.30: organic. Between the two: a lead with at least three indicators, otherwise unclear. The comparison uses the probability once coordination is calibrated, and the score until then.
06
When it cannot finish
If finding communities runs past its time limit, the cycle falls back to scoring the topic as one group and records the error, and the previous report stays visible until a new one lands.
A finding can be reproduced
The method is stored with every finding: which ways of acting together were used, the window for each, the filter and its cut-off, the thresholds, and the version of the installation’s settings in force. A later change of settings does not change what the finding says it used.
The members of each group and the pairs that join them can be exported. A run from the stored method and the exported pairs gives the same report.
Each group also carries checks on each member, for the reviewer: an account created inside the group’s creation window, a profile picture found elsewhere by reverse image search, posting hours against the stated location, posts outside the campaign, and the share of replies that stay inside the group. They are data for the reviewer, never a decider.
What it does not establish
Who is behind the accounts. A finding shows that accounts acted together, far more often than chance. It names accounts, not the people or the organisation directing them.
That the accounts are automated. Campaigns in India have been run by real, long-standing accounts posting templates on a schedule, recruited on WhatsApp; the method looks at behaviour, not at whether an account is a person.
That what they posted is false. Coordination says how something spread, not whether it is true.
Anything organised off the public platforms. Private groups and WhatsApp are not read, so a campaign is seen only once it reaches public posts. The answers AI assistants give are kept out of coordination entirely.
A probability, before calibration. Until coordination has been calibrated on at least 50 topics labelled by two of your analysts, the score is shown as a score.
Bring a push your team suspected was organised, and go through the method that would judge it, step by step.
Request a demo/methodology/coordination · How coordination is judged