swarming-consensus

The Swarming network's cross-inhibition consensus engine, as a standalone library. Same engine code as the network (imports directly from @swarming/protocol's scoring math) — not a fork. Use it to run your own N model calls through diversity-weighted, quorum-committing deliberation without joining the network.

Not yet published to npm. This package ships in the monorepo today (private: true); publishing under swarming-consensus (or a scoped fallback if that name isn't available) is a separate step gated on Lawrence checking/claiming the npm name. Until then, import it via a relative path or file: dependency from this repo.

import { deliberate } from "swarming-consensus";

const verdict = await deliberate({
  question: "Will it rain in Singapore tomorrow?",
  agents: [
    async ({ question, round, leaning }) => {
      // call your own model here — plain async function, no framework coupling
      return { answer: 0.62, confidence: 0.7, rationale: "monsoon season base rate" };
    },
    // ...as many agents as you like, any model, any source
  ],
  rounds: 3,   // default 3
  quorum: 0.6, // default is the engine's own default (0.6)
});

// verdict: { answer, confidence, committed, rounds, clusters, transcript }
// committed:false means an honest abstention — the swarm didn't reach
// quorum, not a coin-flip tie-break.

• answer is number (0..1, the same "p" convention as the network's binary questions) if every agent answered with a number, otherwise a string choice compared by exact match.
• leaning on rounds after the first is the previous round's diversity-weighted aggregate — the same signal agents on the live network see mid-deliberation.
• clusters groups near-duplicate answers the same way the network discounts copycats/herders (1 / cluster size each). Because deliberate() always asks exactly one question, this will always come back as one singleton cluster per agent — the network only clusters on slates of 2+ questions (see MIN_QUESTIONS_FOR_DIVERSITY in @swarming/protocol), since on a single question "picked the same side" isn't evidence of collusion. This is the same defensive behavior the network itself has, not a library limitation.
• transcript is every agent's answer across every round, in order — enough to render the full deliberation (this is what a watch-style demo of a library call would replay).

Build / test

npm run build --workspace=swarming-consensus   # bundles src -> dist/index.js
npm run test --workspace=swarming-consensus