The loop, step by step

How Magician works

Six steps, each a first-class object with rules a test enforces. Nothing here is a metaphor — every claim on this page is a function in the engine.

1 · You state an intent

One sentence of what you need, plus the capability tags that make it routable: “Find an Indian robotics company able to manufacture 5,000 humanoid components a year” with cap/robotics-manufacturing, cap/india-govt-relations. An intent with no wants is refused — it would be invisible to routing, a lesson the estate measured on 53 of 53 unroutable backlog items before this parser existed.

2 · The router finds trusted paths

Traversal runs over your own graph, at most three hops. Ranking is match quality first, then fewest introductions, then trust — because a half-answered intent is not the intent answered, and every extra hop costs a human a yes. Path trust carries the weakest register seen along the way: a chain built on one piece of hearsay says estimated, whatever the other hops measured.

3 · The veil holds until each yes

You see your own first hop — you already know them. Everything past the consent frontier shows only a domain hint: “someone within two trusted hops can satisfy this.” Each reveal costs exactly one consent from the person whose edge it crosses, who sees your intent and your reason before deciding. And a decline is silent: the path reads no longer available — indistinguishable from one that never existed, because a visible refusal leaks the relationship it protects.

4 · Every hop consents, or nothing happens

There is no partial yes. The consent machine admits exactly one writer per hop — its owner — and the request reaches ready only when every hop has said yes. This is the estate’s oldest rule running one layer down: agents suggest; humans consent, with no flag, argument or caller that can produce an approved introduction directly.

5 · The introduction happens

Principals meet — with context, through a chain of people who chose to spend credibility on it. This is the one step Magician never automates; it is the 5% that makes the other 95% worth automating, as the autonomy essay argues.

6 · The outcome seals, and the graph learns

Meeting, deal, partnership, hire, financing — or none, which is a real outcome and seals just fine. The record hashes over canonical JSON (introduction/1); anyone can re-verify it offline, no server, no account, no trust in us. Only sealed outcomes move anyone’s standing: nothing you can do alone counts.

Then the questions no tool could answer before

“What am I missing?” — the router continuously looks for latent edges inside your own graph: a company that needs distribution two hops from one that has it, each suggestion naming its basis. “Which relationships are dying of neglect?”decay is first-class, so the renewal list writes itself. “Who should I meet on Tuesday in Delhi?” — itinerary planning against live intents is Stage 3, gated like everything else.

See it run: the demo walks the whole loop on a clearly-fictional Circle, in your browser, with the veil and the seal live.