Most numbers on a desk are derived. A spread is a difference, a crack is a weighted combination, an attribution is a regression, a days-of-cover figure is a ratio. Each of those takes inputs with their own provenance and produces something new, and the question is what provenance the output carries.
The rule is the floor, never the ceiling. A derived value inherits the worst retrieval mode and the worst reliability across its inputs. One CACHED leg makes the result CACHED. One `official` leg makes an otherwise settlement-grade calculation `official`. Exact arithmetic does not upgrade the weakest observation, and the intuition that a careful calculation improves its inputs is precisely backwards.
There is a judgement the floor rule does not make for you. Pure arithmetic on equally official series can honestly keep that tier. A regression or a cross-instrument comparison should not, even when its inputs are all official, because the output now depends on a fitted relationship that is itself uncertain. The floor is the ceiling on what you may claim; choosing to go lower is the caller's responsibility.
The other half of honest composition is how to degrade. When a leg is unavailable, the options are to omit it silently, to report a zero for it, or to report the result with the missing leg named. The first two both fabricate: a silent omission changes the meaning of the total without saying so, and a zero contribution is indistinguishable from a real measurement of zero. Naming the dropped leg is the only one that leaves the reader able to judge.