Most accounts that get flagged never sent spam. They just fired eleven replies inside one minute.
X runs an internal classifier you cannot query, and it acts on your reach. Public detectors return a score X never sees.
Botometer and its peers have no enforcement power at all. Losing reach is the real danger, and it shows up in no score.
Humans reply in bursts, then vanish for forty minutes. A script firing every 60.0 seconds makes a pattern no human ever has.
Total actions per day and actions per 15 minutes are different numbers. You can clear a daily limit and still look automated for 90 seconds.
Classifiers aggregate. Three weak signals stacking on one account will do what none of them does alone.
Replying to people who mentioned you looks like conversation. Replying to strangers at scale is exactly what detectors are trained to find.
Behavior signals reset in days. Follower structure does not, which is the strongest argument against buying followers.
Search your handle logged out. Check the gaps between your last 200 actions. Count replies to strangers. Divide followers by following.
All seven signals, what each one measures, and the four checks that catch a problem before X does.