signal-fuse

Ask 3 models, get 1 truth.

BFT-inspired multi-LLM consensus. Query multiple AI models in parallel, reject outliers, return a single confidence-scored answer.

You ask GPT-4o a question. It says "yes." You ask Claude the same question. It says "no." Who's right?

Most people pick one model and trust it. That's like driving with one eye closed.

signal-fuse queries N models in parallel and uses Byzantine fault tolerance (the same math behind distributed consensus systems) to find agreement. Outlier responses are flagged. You get one answer with a confidence score.

$ signal-fuse "What is the melting point of tungsten?"

  signal-fuse v1.0.0
  ──────────────────────────────────────
  prompt    What is the melting point of tungsten?
  models    openai, anthropic, gemini

  ✔ 3/3 models responded

Tungsten has a melting point of 3,422°C (6,192°F), making it
the highest melting point of any pure metal.

  Confidence: ██████████████████████████████ 98%
  Method:     unanimous (3/3 models)
  Latency:    1243ms slowest
  ──────────────────────────────────────

# Global
npm install -g signal-fuse

# Or run directly
npx signal-fuse "your question"

Set at least 2 API keys as environment variables:

export OPENAI_API_KEY="sk-..."
export ANTHROPIC_API_KEY="sk-ant-..."
export GEMINI_API_KEY="..."

# Auto-detect available providers
signal-fuse "What is the capital of France?"

# Specific models
signal-fuse -m openai,anthropic,gemini "Explain quantum entanglement"

# Verbose — see individual responses + outlier analysis
signal-fuse -v "Should I use microservices or a monolith?"

# JSON output (for piping to jq)
signal-fuse -f json "Is water wet?" | jq '.confidence'

# Strict consensus (80%+ similarity required)
signal-fuse -t 0.8 "Complex nuanced question..."

# Quiet mode — just the answer, no metadata
signal-fuse -q "What is 2+2?"

Your prompt
    │
    ▼
┌─────────────────────────────────────┐
│  Parallel Query (native fetch)      │
│  ┌───────┐ ┌───────┐ ┌───────┐     │
│  │ OpenAI │ │Anthro.│ │Gemini │     │
│  └───┬───┘ └───┬───┘ └───┬───┘     │
└──────┼─────────┼─────────┼──────────┘
       │         │         │
       ▼         ▼         ▼
┌─────────────────────────────────────┐
│  Normalize → Bigrams → Jaccard     │
│  Pairwise similarity matrix        │
└─────────────────────────────────────┘
       │
       ▼
┌─────────────────────────────────────┐
│  Majority clustering                │
│  Pick largest agreement group       │
│  Reject outliers                    │
└─────────────────────────────────────┘
       │
       ▼
  Consensus + Confidence score

1. Parallel query — All providers queried simultaneously via fetch
2. Text normalization — Lowercase, strip punctuation, collapse whitespace
3. Bigram extraction — Word-level bigrams create a fingerprint of each response
4. Jaccard similarity — Pairwise similarity between all response fingerprints
5. Majority clustering — Response with the most similar neighbors wins
6. Confidence = agreement_ratio × average_similarity

This is adapted from the Byzantine fault tolerance algorithm used in Aegis for drone swarm sensor fusion — same math, different domain.

import { queryAll, fuse } from "signal-fuse";

const responses = await queryAll("What is the capital of France?", ["openai", "anthropic"]);
const result = fuse(responses, 0.6);

console.log(result.consensus);   // The agreed answer
console.log(result.confidence);  // 0-1 confidence score
console.log(result.method);      // "unanimous" | "majority" | "best-effort"
console.log(result.outliers);    // Responses that disagreed

Jaccard similarity on word bigrams:

• Tokenize response into word pairs: "the capital of france" → ["the_capital", "capital_of", "of_france"]
• Similarity = |intersection| / |union| of two bigram sets
• Fast, language-agnostic, no external dependencies

Why not embeddings?

• Embeddings need an API call per response (cost, latency)
• Bigram Jaccard is zero-dependency, runs in <1ms, and works surprisingly well for factual agreement
• For production use, you could swap in embeddings — the consensus logic is provider-agnostic

Issues and PRs welcome. This is a small, focused tool — keep it that way.

git clone https://github.com/Vitalcheffe/signal-fuse.git
cd signal-fuse
npm install
npm test

MIT