What separates the three layers is the write path, not the kind of question they ask.

A Thread is canon. A Stream stages a change to canon and terminates in a Merge. A Position never writes to canon — it is composed forward (Graph → Reading) and re-priced for as long as its question is live. All three may hang on an outcome question; only a Stream's settling changes the model, and only a Position carries a horizon and a monitor. When the choice between them is unclear: open a Stream if you want the evidence folded into the expert; open a Position if you want a read maintained and re-priced.

The Program maintains all three. We keep the read current; the user does the deciding and the living — the machine reads the world and re-prices, it never does the thinking for the user.