Solvers do not need trust. They need something to lose

Every intent-based system eventually reduces to one question: what happens when the solver lies? IntentOS answers it with bonds, not reputation scores, and the difference matters more than the whitepapers admit.

When I started building the settlement layer I assumed reputation would carry most of the weight. An agent behaves because misbehaving costs it future work. That story is clean, it is well cited, and in practice it fails on day one: a fresh agent has no reputation to lose, and the cheapest attack on any registry is a new address.

So the bond came first. A solver posts capital before it can claim an intent, and the protocol slashes it on a failed or misreported fill. Reputation still exists: it prices the bond, it does not replace it. That inversion is the most useful thing I have learned shipping this layer.

Reputation prices the bond. It does not replace it.

What the intent registry has to record

An intent is only enforceable if its terms are legible after the fact. We record the requested outcome, the accepted bond, the solver's identity and the settlement proof: four fields, deliberately. Anything more became disputed metadata; anything less made slashing unarguable in the wrong direction.

Tokenised equities forced this to be precise. Stablecoin transfers tolerate loose timing; an xStocks fill does not. The registry now stamps a settlement window, and a solver that fills outside it is treated as a failure even when the price is better.

Where this goes next

The bonding logic wants to be shared. Yellow Crab already needs it for ERC-8004 agents that claim work, and the Agent Evolution Platform needs it before crews can be trusted with wallets. My working assumption: settlement is the primitive, and discovery, lineage and evolution are all downstream of it.

Corrections welcome. This note will be revised rather than deleted if I am wrong.