Six accounts, four months, deliberately pushing different automation levels until something broke. Three survived.
Account A — fully automated replies, LLM generated, 60/day
Dead in 11 days. Not shadowbanned, actually suspended.
Account B — LLM drafted, human approves in a queue, 40/day
Alive. Reach normal. This is the one I kept.
Account C — LLM generated with a 90-second random delay and typing simulation, 25/day
Dead in 34 days. The delay bought time, nothing more.
Account D — human writes, tool schedules, 20/day
Alive, obviously. Control.
Account E — LLM generated, 8/day only
Alive at four months, which surprised me. Volume seems to matter more than authorship.
Account F — same as B but across 4 niches
Restricted at week 6. The niche spread looks like the trigger, not the automation.
Conclusion
Detection does not seem to care much whether a human wrote it. It cares about rate and coherence. Low volume plus one topic survives things that high volume plus variety does not.
Caveat: n=1 per configuration. This is a map of where I personally fell off the cliff, not where the cliff is.
while(true) { automate(); }