What is the CommunityOS scoring methodology?

A two-stage deterministic pipeline. Stage one, Bot-Kill, filters inauthentic and inactive accounts using four signals: an activity floor, follower-to-following imbalance, posting-burst patterns, and linguistic uniformity. Stage two scores every surviving account across four archetypes (Champion, Amplifier, Builder, Early Adopter) using a fixed model weighted 60 percent on linguistic features and 40 percent on normalized vanity metrics. The same inputs always produce the same outputs.

Why is the scoring deterministic instead of using an LLM?

Because the scores drive outreach and rewards, and those decisions must be reproducible, auditable, cheap to run at scale, and free of drift from third-party model updates. A rules-based engine gives the same answer twice and can explain any score feature by feature; an LLM classifier can do neither reliably.

Is the CommunityOS methodology open?

Yes. This methodology paper is published under CC-BY-SA 4.0, so it may be quoted, cited, and adapted with attribution. The specific feature list, weights, and a held-out validation sample are released as a companion appendix so the scoring can be reproduced.

Methodology — Deterministic community scoring

Abstract

CommunityOS classifies the followers of an X (Twitter) account to identify which real people are worth activating, and in what way. The method is deliberately deterministic: a fixed filtering stage followed by a fixed scoring stage, with no language-model inference in the scoring path. This paper documents the two stages, the reasoning behind the 60/40 linguistic-to-metric weighting, the four-archetype target, the validation from the first production deployment, and the method's stated limitations. It is published under CC-BY-SA 4.0 so it can be quoted, cited, and reproduced.

1. The problem

A follower list is not a community. It is a mixture of real community and noise — bots, farms, and abandoned accounts — in unknown proportions. Ranking that list by follower count surfaces exactly the wrong accounts, and any metric computed over the unfiltered base (engagement rate, reach) inherits the contamination. The task is therefore two problems, not one: separate real accounts from noise, then rank the real accounts by how they can move for the brand. Both must be solved in a way that is reproducible and auditable, because the output drives outreach and, downstream, rewards.

2. Stage one — Bot-Kill

Before any scoring, every account passes through a filter we call Bot-Kill. It removes accounts that fail authenticity or activity screening, using four deterministic signals designed so that evading all four simultaneously costs more than a farm earns:

Activity floor. A minimum of genuine, recent activity. Abandoned and dormant accounts are unreachable for activation regardless of intent, and are removed first.
Follower-to-following imbalance. Normalized ratio and growth-shape checks that separate organic accumulation from follow-back harvesting and follow-for-follow rings.
Posting-burst patterns. Timing analysis that detects mechanical cadence and cross-account synchronization — the fingerprint of one operator running many identities.
Linguistic uniformity. Detection of templated or generated text: recycled phrasings across accounts, unnaturally consistent structure, engagement-bait grammar.

The signals are individually evadable and jointly expensive to beat, which is the design goal: not a filter no bot can pass, but a filter no bot can pass profitably. Bot-Kill is deliberately conservative — it prefers to hold back a quiet real account over admitting a farm account to an activation campaign, because the downstream cost of the second error is higher.

3. Stage two — archetype scoring

Each account that survives Bot-Kill is scored across four archetypes. The score combines two families of features with fixed weights: 60 percent linguistic features (depth, topical focus, originality, on-domain specificity) and 40 percent normalized vanity metrics (follower, engagement, and activity measures, normalized so genuine smaller accounts are not dominated by inflated larger ones). The weighting encodes a single editorial claim: how an account writes is a better predictor of conviction than how large it is, and language is harder to fake at scale than a follower number.

The scoring uses no sampling, no temperature, and no prompt. Run the same account through twice and the score is identical; run it next quarter on unchanged data and it is still identical.

4. The four archetypes

The target set is four because four is the number at which every class still binds to a distinct action. Each archetype answers the operator's real question — what should I ask of this person?

Archetype	Dominant currency	The ask
Champion	Conviction	Early, substantive information worth their voice
Amplifier	Reach	Effortless distribution at the right moment
Builder	Substance	Access, materials, recognition for what they ship
Early Adopter	Timing	First access and a short feedback loop

Three classes would merge groups that need opposite treatment; five reliably introduces a class that maps onto an existing action or onto accounts that should not receive outreach at all. Membership is not fixed — accounts move between archetypes over time, and those transitions are themselves signal.

5. Output — the ranked queue

The scored population is ordered into a ranked queue that answers one question: who deserves attention this week, and why. Recency, relevance, and archetype combine so that a recently active, on-topic account outranks a larger but dormant one. The output is a short action list, not a dashboard — designed to be cleared by a human operator before it goes stale.

6. Validation — first production deployment

The method's first production run scanned the full follower base of one Web3 project (Mintlayer):

78,181 followers scanned.
90.96 percent filtered by Bot-Kill as inauthentic or inactive.
5,806 real community members ranked across the four archetypes.
298 surfaced as immediate-action in the ranked queue.

The 90.96 percent filter rate is one project's result, not a universal constant — audiences that have not run incentive campaigns filter substantially lower, and established consumer brands lower still. It is reported here as a measured outcome of a specific deployment, not a claim about follower lists in general.

7. Reproducibility

Because both stages are deterministic, any account's result can be recomputed and traced to the exact signal or feature that produced it. To make the method reproducible by third parties, CommunityOS releases a companion appendix — the specific feature list, the weights within the 60/40 split, and a held-out validation sample — so the scoring can be reproduced on independent data. This paper describes the method; the appendix pins the parameters.

8. Limitations

Scope. The method targets X (Twitter) audiences and the activation problem. It is not social listening, not a quest platform, and not pipeline scoring.
Conservatism. Bot-Kill's bias toward filtering means some quiet real accounts are held back. This is an intentional trade, not an accident.
No universal rates. Filter and archetype-mix figures are deployment-specific and should not be generalized across audiences.
Language dependence. The 60 percent linguistic weighting assumes sufficient text to analyze; extremely low-post real accounts carry more uncertainty and are treated accordingly.

License and citation

This methodology paper is published under the Creative Commons Attribution-ShareAlike 4.0 International license. You may share and adapt it, including for commercial purposes, with attribution and under the same license. Suggested citation: CommunityOS, "Deterministic community archetype scoring," methodology paper v1.0, Levee Investments LLC, 2026. https://communityos.so/methodology.

Related reading: the engine overview, the Bot-Kill teardown, the case for deterministic over LLM scoring, the four-archetype design, and the first production scan in full.