CommunityOS does not run on a large language model. The scoring engine is rules-based, reproducible, and transparent — built so two runs over the same input produce the same archetype score, every time. This page explains how, in enough detail to be useful and not so much that the methodology becomes the moat.
A large language model can read a bio and a post history and tell you, in plausible prose, that someone is a Champion or a Builder. The output reads well. It also changes between runs, sometimes for the same input, often for inputs that differ in ways the prompt was never told mattered. Two operators looking at the same model output disagree about what to do next.
For a feature that classifies who gets activated, who gets ignored, and who gets paid, that level of stochasticity is unacceptable. CommunityOS is built on the opposite premise: the scoring engine is deterministic. Given the same input, the same parameters, and the same model version, the engine produces the same archetype score every time. That property is not a marketing claim. It is what allows every downstream surface — the queue, the report, the rewards — to be defensible.
The single most-important design choice in the engine is that what someone says about you matters more than how many people heard them.
Three families of signal, each contributing to the linguistic component of the score:
Vanity metrics are not useless. They are misleading when used as the only signal. The engine normalizes them in two ways:
The most common failure mode in community scoring is to apply the archetype model to a follower list that contains a substantial percentage of bots, farmers, and inauthentic accounts. The model will produce scores. The scores will be garbage. The campaign will get built on top of the garbage.
Bot-Kill runs first. Before the 60/40 model touches anything, the candidate pool is filtered through a set of behavioral, structural, and engagement-pattern checks that remove the obvious cases. In the Mintlayer pilot, roughly 64 percent of the raw follower count was filtered at this stage — typical for Web3 audiences, lower for established brand accounts.
The filter does not require perfect bot detection. It requires high precision at the cost of recall. A real Champion accidentally filtered out can be added back on review. A farmer that survives the filter pollutes every downstream surface. So the filter is tuned to be aggressive, with explicit append paths for false positives.
Every follower in a scan receives a four-dimensional score: Champion, Amplifier, Builder, Early Adopter. The four scores are independent. A single person can rank high in two or three archetypes simultaneously — Champions often have Amplifier reach, Builders often start as Early Adopters.
For each candidate: bio text, last 200 posts (mixture of original posts and replies), follower count, following count, account age, recent engagement velocity, and a graph signal indicating whether the candidate appears in the project’s own engagement history.
The full methodology paper, including the exact weight matrix and the seed-corpus extraction rules, is in development. It will be published under CC-BY-SA so the methodology can be cited, reproduced, and challenged. The publication date is set for Q3 2026.
Pilot scans include a 30-minute methodology walk-through with Gai, the engine’s author.
Request a methodology call