[Research] ALIGN primitive: Formalizing agreement detection

## Research Question

INTENT proposes an **ALIGN** primitive—an explicit checkpoint where sender and receiver confirm shared semantic frame before proceeding.

Can we formalize when ALIGN is needed, when it succeeds, and when it fails?

## The ALIGN Protocol

```
Sender: Intent { term: "efficient", meaning: ??? }

Receiver: ALIGN {
term: "efficient",
my-interpretation: "minimize-wall-clock-time",
is-this-your-meaning?
}

Sender: ALIGN-RESPONSE {
confirmed: false,
divergence: significant,
action: CLARIFY { ... }
}
```

## Formalization Goals

1. **Detection**: When is ALIGN triggered? (ambiguity threshold? domain mismatch? calibration gap?)
2. **Success criteria**: What constitutes successful alignment? (exact match? acceptable overlap? human escalation?)
3. **Failure modes**: What happens when ALIGN fails repeatedly? (abort? escalate? partial agreement?)

## Connection to Other Work

- Grounding in communication (Clark & Brennan)
- Consensus protocols in distributed systems
- Alignment problem in AI safety

## Research Direction

Formalize ALIGN as a state machine with:
- States: proposing, querying, confirming, diverging, resolved
- Transitions: triggered by ambiguity, calibration gaps, or explicit request
- Termination: success (proceed), failure (escalate), or timeout (abort)

---
*Research thread: Explicit alignment in negotiation protocols*