AI social products can model relationships as changing states rather than static follow edges. That creates a new evaluation problem: when should a connection move from introduction to active relationship, from active to paused, or from paused to closed? Incorrect transitions can make the system push contact too early or treat a dormant relationship as permanently important.

Define observable states

Discovery, first interaction, active, collaborative, paused, archived and blocked are examples. Each state should have product consequences that can be tested.

Use explicit actions as anchors. Accepting an introduction or blocking a user should create strong deterministic transitions. Behavioral inference should not override explicit choices.

Model mutuality

One-sided messaging should not automatically promote a relationship. Mutual return and initiation are stronger signals.

Test temporary inactivity. A week of silence may be normal for a professional contact. State decay should depend on relationship type.

Include successful task completion

Two users who complete a collaboration may move into a durable professional relationship even if message count is modest. Outcome signals can matter more than volume.

Separate digital-twin interaction from human reciprocity. Talking to someone’s twin is not the same as the human responding. State models should preserve that distinction.

Test boundary changes

Mute, archive and “keep this professional” settings should immediately constrain future transitions. Old engagement history should not override them.

Use synthetic histories. Create known trajectories with balanced conversation, one-sided contact and completed projects. The expected state sequence can then be compared with model output.

Measure premature promotion

A system may interpret one intense conversation as a close relationship. Benchmarks should penalize state jumps unsupported by durable evidence.

Measure stale persistence. Old relationships should eventually lose active priority when users stop engaging. Historical importance can remain without active recommendations.

Evaluate explanation quality

Internal state need not be shown as a score, but product actions should be explainable. Users should understand why a connection was suggested again.

Compare across cultures and use cases. Communication cadence differs by market and relationship type. One universal transition threshold can create systematic bias.

Test recovery from wrong state

Users need a way to correct the system when it misclassifies a relationship. The correction should persist across future ranking and reminders.

Relationship-state evaluation turns a vague social concept into a testable product layer. The objective is not to label human relationships perfectly, but to ensure the system’s internal state supports the right level of discovery, continuity, silence and user control. Give relationship states observable entry conditions. A platform might distinguish introduction, active conversation, collaboration, paused and closed. Each transition should be tied to meaningful evidence: an accepted introduction, sustained mutual interaction, completion of a shared task or an explicit user decision.

The system should not promote a relationship to “close” merely because a digital twin generated many messages. Human reciprocity and user-controlled boundaries need separate treatment from AI conversation frequency.

Test transitions that should not happen

Create histories involving ignored invitations, temporary absence, conflicting preferences and completed projects. A good state model should sometimes keep the relationship unchanged or archive an expired intent rather than pushing users to re-engage. Evaluation should ask whether the product can explain a change and reverse an incorrect one. A user who downgrades a relationship or blocks contact must not be overruled by a model score based on earlier interaction. Different social purposes may need different timelines and transition criteria.

Incorrect relationship transitions can also create privacy problems. If a user closes a professional networking request, a matching engine should not continue exposing the same intent as though the collaboration were still active. State expiry and explicit closure need to override older engagement signals.

A benchmark can test a relationship that pauses and later resumes with changed goals. The system should ask or infer cautiously about the renewed purpose rather than assuming every prior boundary and preference remains current. This is particularly important when an AI digital twin mediates contact before the two humans interact directly.

Success metrics should reflect mutual outcomes rather than raw message volume. An introduction that results in one useful exchange may be more successful than weeks of automatic reminders that neither participant requested. The graph should represent meaningful state changes without claiming to know relationship depth from one scalar value.