a tie is a contract, not a feature

The obvious build is a relationship service: a table of who is bound to whom, a query surface, a component that can answer membership questions on behalf of everything else. The record refuses it. Ties between identities are agreements made end to end between the parties, and no component has the job of knowing about them.

The tension is that a platform-visible graph is what makes coordination cheap. Find a common counterparty, check whether a group's decision binds you, route a request through a connection you already hold — all of it is trivial with a shared table and none of it is trivial without one. An end-to-end agreement makes every one of those questions the parties' own problem.

The decision keeps the platform ignorant on purpose. It records that an agreement exists; the members, the terms, and the shape of the graph stay with the people who made it. Anything that needs the graph has to be built by somebody who owns it, and that builder is admitted like any other participant.

What it costs is the absence of a global answer to who is connected to whom, duplicated work as several tools each construct the part of the graph they need, and no shared arbitration when two parties disagree about whether a membership ever happened.

What it proves is that the team could see the subsystem everyone would ask for, priced what it would centralise, and declined to make a relationship registry the place every participant must pass through. That refusal is the difference between a platform and a landlord.