@relaylaunch/councilverse-voting

Three-valued voting for AI council debates. Agents cast KEEP, REFUSE, or ABSTAIN instead of forced binary positions. Verdicts determined by argument quality, not headcount.

Based on the CAMP paper (arXiv:2604.00085).

Install

npm install @relaylaunch/councilverse-voting

Usage

import { buildAgentVote, tallyVotes } from '@relaylaunch/councilverse-voting';

// Build votes from agent response text
const votes = [
  buildAgentVote('agent-1', 'Strategist', 'council-1', agentResponse1),
  buildAgentVote('agent-2', 'Risk Analyst', 'council-1', agentResponse2),
  buildAgentVote('agent-3', 'Devil\'s Advocate', 'council-1', agentResponse3),
];

// Tally using quality-weighted scoring
const result = tallyVotes(votes);

console.log(result.outcome);  // "keep" | "refuse" | "deadlock"
console.log(result.method);   // "strong_consensus" | "quality_weighted" | "lead_fallback"
console.log(result.margin);   // 0.0 - 1.0

How It Works

Three-Valued Votes

• KEEP -- agent supports the proposal
• REFUSE -- agent opposes the proposal
• ABSTAIN -- agent signals "outside my expertise"

ABSTAIN prevents noise from agents opining on domains they don't understand.

Quality-Weighted Scoring

Verdicts are determined by argument quality, not simple headcount:

1. Strong consensus -- all active voters agree (2+ agents)
2. Quality weighted -- weight = argument_quality * confidence, margin >= 15%
3. Lead fallback -- deadlock when margin < 15%

Argument Quality Scoring

Arguments scored on evidence, structure, and specificity:

• Evidence markers ("data shows", "according to", "based on")
• Structure markers ("first", "second", "therefore", "in contrast")
• Specificity markers (percentages, dollar amounts, dates)

API

| Function | Description | |----------|-------------| | buildAgentVote(agentId, agentName, councilId, text) | Build vote from raw agent text | | tallyVotes(votes) | Quality-weighted vote tally | | detectVote(text) | Extract KEEP/REFUSE/ABSTAIN from text | | detectConfidence(text) | Score confidence level (0.1-1.0) | | scoreArgumentQuality(text) | Score argument quality (0.05-1.0) |

License

MIT -- RelayLaunch LLC