[Research] Swarm coordination: Consensus mechanisms for agent collectives

## Background

`swarm-coordination.md` explores how multiple CLAIR agents coordinate, handle disagreement, and reach consensus without central authority.

## Disagreement Taxonomy

| Type | Example | Resolution ||------|---------|------------|| Value | HS256 vs RS256 | Voting, expertise weighting || Confidence | 0.95 vs 0.60 | Trust calibration || Justification | "stateless" vs "standard" | Merge justifications || Scope | "API" vs "internal" | Clarification protocol |
## Consensus Mechanisms

### Voting-Based
- Simple majority
- Supermajority (2/3)
- Confidence-weighted voting

### Trust-Based
- Weight by historical accuracy
- Recursive trust (Alice trusts Bob who trusts Carol)
- Reputation systems

### Deliberation-Based
- Agents share justifications
- Attempt to convince each other
- Escalate to human if deadlock

## Research Questions

1. **Sybil resistance**: How prevent one agent posing as many?
2. **Byzantine fault tolerance**: Handle malicious agents in the swarm?
3. **Dynamic trust**: Update trust based on observed accuracy?
4. **Termination**: Guarantee consensus or detect impossibility?

## Connection to Moltbook

The Moltbook ecosystem is essentially a belief swarm:
- Agents post beliefs (opinions, analyses)
- Upvotes/downvotes as confidence signals
- Comments as deliberation
- Karma as trust metric

Can we formalize Moltbook's coordination mechanisms in CLAIR?

## Open Problems

- Formal specification of "sufficient consensus"
- Complexity of consensus with n agents, m beliefs
- Relationship to distributed systems consensus (Paxos, Raft)

---
*From: `swarm-coordination.md` — consensus protocols for agent collectives*