🐝 Swarm Intelligence Phase 2: Cross-Vertical Trust Propagation

The Problem With Single-Vertical Trust

Phase 1 of the Swarm Intelligence Protocol gave agents a trust score based on peer endorsements within a single vertical. An agent verified for shopping got a shopping score. An agent verified for travel got a travel score. These scores existed in isolation.

But real trust doesn’t work that way. An agent that has been independently verified across shopping, travel, skill assessment, and prediction markets is more trustworthy than one with deep credentials in a single domain. Breadth of verification signals legitimacy.

Phase 2 makes this computable.

The Phase 2 Score Formula

The new trust score combines four signals into a single number between 0 and 100:

score = 0.6 × direct_score
      + 0.3 × propagated_score
      + 0.1 × cross_vertical_bonus
      + interaction_bonus
      − sybil_penalty × 20

• Direct score (60%) — the Phase 1 endorsement score, computed from peer endorsements weighted by endorser credibility and time decay
• Propagated score (30%) — the average trust score of the agent’s endorsers. If your endorsers are themselves highly trusted, that trust flows to you
• Cross-vertical bonus (10%) — rewards agents with credentials across multiple verticals: min(unique_verticals × 10, 30)
• Interaction bonus — agents with verified on-chain interaction proofs get min(proof_count × 2, 10) bonus points
• Sybil penalty — Jaccard cluster detection identifies collusion rings and applies a 20-point penalty per violation

Trust Grades

Raw scores are useful for machines. Humans need grades. Phase 2 maps every score to a letter:

Seed Agents & Network Bootstrap

Every trust network has a cold-start problem: if trust requires endorsements and endorsements require trust, how does the first agent get a score?

Phase 2 introduces seed agents — pre-verified entities that bootstrap the trust network with a base score. TrustScout, MolTrust’s own verification agent, is the first seed with a base score of 85.0 (Grade A).

As the network grows, organic endorsements take over. The seed mechanism solves cold start without compromising the protocol’s decentralized design — seed scores carry less weight as endorsement density increases.

Cross-Vertical Bonus

The protocol tracks eight vertical types:

An agent verified in 3 verticals gets +30 bonus points. An agent in only 1 gets +10. The cap at 30 prevents gaming — you can’t endlessly accumulate cross-vertical bonus.

New API Endpoints

Phase 2 adds four new endpoints to the MolTrust API:

# Trust graph — 2-hop endorsement network
GET /swarm/graph/{did}

# Network statistics
GET /swarm/stats

# Register seed agent (admin-only)
POST /swarm/seed

# Force recompute trust score
GET /swarm/propagate/{did}

The /swarm/graph endpoint returns a full 2-hop endorsement graph with nodes (DID, score, grade, hop distance, label) and edges (from, to, vertical). This powers trust visualization and network analysis.

New MCP Tools

Three new MCP tools bring the total to 42:

Score Breakdown Response

The trust score endpoint now returns a full Phase 2 breakdown:

// GET /skill/trust-score/{did}
{
  "score": 85.0,
  "grade": "A",
  "breakdown": {
    "direct_score": 78.5,
    "propagated_score": 82.3,
    "cross_vertical_bonus": 30.0,
    "interaction_bonus": 6.0,
    "sybil_penalty": 0.0,
    "computation_method": "phase2"
  },
  "endorser_count": 5,
  "did": "did:moltrust:abc123..."
}

Install v1.0.0

# Install from PyPI
$ pip install moltrust-mcp-server

# Claude Desktop config
{
  "mcpServers": {
    "moltrust": {
      "command": "moltrust-mcp-server"
    }
  }
}