OYOTTA records two kinds of learning. They must not be collapsed.

Information-state learning

A later decision may change because instrumentation became live, private aggregation became machine-consumable, expected information gain fell, or a duplication penalty rose. That is a change in what the system knows about its own apparatus. It is not evidence that visitors responded.

Behavioural-response learning

A ranking or allocation change from visitor behaviour requires paired exposure and response under one allocation mode, with enough denominator that a posterior interval can separate two objects. Until that threshold is met the honest result is no ranking change yet.

What this note does not claim

Public empty experiment buckets are a privacy boundary, not observed zeros. A successful URL ping is not indexation. OYOTTA events already exist on owned surfaces; they are not converted into a Stream re-rank here.

Listen, Visual, Editorial, Watch, Discover, Archive, Collections, and Interactive keep being built while behavioural ranking is held.

The public Graph explorer is a filtered public-safe projection. It is not the Stream corpus and it is not the operational object graph. Different layer counts are expected when their filter policies differ. A later ranking change from visitor behaviour would be recorded separately as behavioural-response learning; that record does not exist yet.