[Research] Multi-agent beliefs: From single-agent traces to collaborative reasoning

## Background

`multi-agent-beliefs.md` formalizes belief attribution and combination across multiple agents (Claude, GPT, human, systems).

## Key Concepts

### Agent-Attributed Beliefs
```clair
belief {
value: verify_token_impl,
confidence: 0.91,
agent: AI("claude", "opus-4"), -- NEW
justification: authored(...)
}
```

### Nested Beliefs (Beliefs about Beliefs)
```clair
-- Alice believes Claude's belief is well-justified
alice_about_claude : Belief>
alice_about_claude = belief {
value: belief_claude,
confidence: 0.85, -- trust in Claude, not code correctness
agent: Human("alice"),
justification: reviewed_justification(...)
}
```

### Belief Combination Strategies
- **Maximum**: optimistic, take highest confidence
- **Minimum**: conservative, take lowest confidence
- **Weighted**: by agent track record
- **Consensus**: structural agreement required

## Research Questions

1. **Trust graphs**: How do we model Alice trusts Claude, Bob trusts Alice, transitive trust?
2. **Epistemic depth limits**: How many levels of `Belief>` are tractable?
3. **Disagreement resolution**: When agents disagree, what protocols resolve it?
4. **Human-in-the-loop**: How does human override factor into multi-agent consensus?

## Connection to INTENT

Multi-agent beliefs are essential for the Thinker/Doer architecture:
- Thinker and Doer are different agents
- Doer may have beliefs about Thinker's reliability
- Human reviewer has beliefs about both

The current CLAIR v1 doesn't distinguish agents—should it?

## Implementation Path

- Extend belief type with `agent` field
- Define belief combination operators
- Model consensus protocols from `swarm-coordination.md`

---
*From: `multi-agent-beliefs.md` — partially developed, not in current spec*