Holding the computation fixed, we vary the compression factor c (macro order or base). Generated-trace positions and cache scale as 1/c, generated-trace attention operations as 1/c2, and the trace information stays flat; a fixed input prefix is not compressed.