In one paragraph

Saphan Studio is a control plane that dispatches AI coding agents to a fleet of machines you own. Its model rests on four decisions. A human decides, always — the engine routes work and generates artifacts; it never approves a merge and never treats a timeout as consent. Nothing is trusted because it says so — machines are admitted by an explicit owner act with a certificate binding, instructions trace to a signing key, and a machine's claim about what it can do carries a provenance grade. The record is the system — governance registries are Ed25519-signed append-only logs, and the human-readable files are projections of them. When a control cannot be enforced, the run does not happen — a missing sandbox facility, an unreadable policy value, an unadmitted machine, or an ambiguous target are all refusals with named reasons, never a silent fallback to a weaker posture.

---