Bot-Kill is the filter that runs before anything else in the CommunityOS engine. Four deterministic signals — an activity floor, follower-to-following imbalance, posting-burst patterns, and linguistic uniformity — decide which accounts never reach scoring at all.
In the first production scan, those four signals filtered 90.96 percent of 78,181 followers. This post opens the filter up: what each signal measures, why farms cannot cheaply evade all four at once, and what Bot-Kill deliberately does not try to do.
Why filter before scoring?
Because scoring garbage produces ranked garbage. An archetype engine pointed at an unfiltered Web3 follower list would dutifully classify tens of thousands of shells — and a farm account with scraped crypto vocabulary and inflated engagement can score deceptively well on any model. The economics only work if inauthentic accounts are removed before the expensive analysis, which is also why Bot-Kill is built from cheap, hard signals rather than deep content analysis. Filter first, think second.
Signal 1 — What is the activity floor?
The bluntest test: does this account exhibit a minimum level of genuine, recent activity? A huge share of any old follower list is not adversarial at all — it is dead. Abandoned accounts, one-week experiments, airdrop wallets with an X handle attached. They will never see a post, never complete a mission, never matter. The activity floor removes them without ceremony, and on aged Web3 audiences it is the single highest-volume filter of the four. Not every filtered account is a bot; every filtered account is unreachable, which for activation purposes is the same thing.
Signal 2 — What does follower-to-following imbalance reveal?
Growth mechanics. Real accounts accumulate followers by being interesting; farm accounts accumulate them by following tens of thousands of accounts and harvesting the follow-backs, or by buying entry into follow-for-follow rings. Both strategies leave a signature in the ratio and in its shape over the account's lifetime that organic growth does not produce. The signal is normalized rather than a naive threshold — a genuine newcomer who follows 400 accounts and has 30 followers is fine; an account following 40,000 with the engagement profile of a stone is not.
Signal 3 — What are posting-burst patterns?
Timing forensics. Human posting is irregular — it clusters around waking hours, events, moods. Automated accounts betray themselves with mechanical cadence: posts at fixed intervals, activity bursts synchronized across whole clusters of accounts (the fingerprint of one operator scheduling a farm), long silences broken by sudden coordinated engagement when a paid campaign fires. Individual accounts can look plausible; the coordination across them cannot. This cluster behavior is the practical face of a Sybil attack — one operator, many identities — and it is the signal farms find hardest to hide, because breaking the synchronization breaks the farm's economics.
Signal 4 — What is linguistic uniformity?
The language test, and the philosophical heart of the engine. Real people write with variance: vocabulary drifts, sentence length wobbles, topics wander off-brief. Farm output is generated or templated, and it shows — recycled phrasings across accounts, unnaturally consistent structure, engagement-bait grammar, replies that parse as topically adjacent but say nothing. This is the same conviction that puts language at 60 percent of the archetype scoring model: how an account writes is harder to fake at scale than what its metrics claim.
Why do the four signals work together?
Because evading all four simultaneously costs more than the farm earns. Beating the activity floor requires sustained posting. Beating the ratio check requires slow organic-shaped growth. Beating burst detection requires desynchronizing the fleet, which destroys the operational efficiency that makes farms profitable. Beating linguistic uniformity requires generating genuinely varied, substantive text across thousands of accounts, indefinitely. Each signal is individually evadable; jointly they push the cost of a convincing fake account toward the cost of just being a real person.
The goal is not a filter no bot can pass. It is a filter no bot can pass profitably.
What does Bot-Kill deliberately not do?
Three scope decisions worth stating plainly. It does not claim a universal bot rate — 90.96 percent was one Web3 project's number, typical for audiences that have been through airdrop cycles and materially higher than an established brand should expect. It does not chase certainty on every account — it is conservative by design, preferring to hold back a quiet real account rather than let a farm into an activation campaign, because the downstream cost of the second error is higher. And it does not hide its reasoning: the filter is deterministic, so any account's fate can be recomputed and traced to the exact signal that fired — the same auditability standard the whole engine is held to.
What survives Bot-Kill goes on to archetype scoring, ranking, and the operator queue — the pipeline described across the engine page and the glossary. In the Mintlayer scan, what survived was 5,806 real people. That is the number everything else is built on.
Quick answers
What is Bot-Kill?
Bot-Kill is the pre-scoring filter in the CommunityOS engine. It removes farm, shell, and inactive accounts from a follower list before archetype scoring, using four deterministic signals: activity floor, follower-to-following imbalance, posting-burst patterns, and linguistic uniformity.
How much of a follower list is typically bots?
It varies widely by audience. In the first CommunityOS production scan of a Web3 project, 90.96% of 78,181 followers were filtered. Audiences that never ran incentive campaigns, and established consumer brands, filter substantially lower.
Can Bot-Kill be wrong about an account?
Yes — it is deliberately conservative, so quiet real accounts can be held back. Because the filter is deterministic, any account's result can be recomputed and traced to the exact signal that fired, and thresholds are tuned from production scans.