Rank Order Under a Capacity Benchmark Is Conditional on the Observation Band and the Delay Horizon

Reservoir-computing benchmarks increasingly report rank orderings, but rarely the observation band or the delay horizon at which the ranking was taken. This note measures both conditions on a common grid, on two architecture families. Under a readout-only change of band, the fastest system loses -6.61 information processing capacity [-6.89, -6.33] while its kernel rank moves -0.50 [-1.27, +0.27]: a readout band destroys memory and leaves separation intact, because rank measures, under the ensemble protocol, are computed from one state per input stream. That insulation belongs to the protocol and not to rank measures: under the field's second protocol the manipulation moves generalisation rank in both, by up to twenty-eight points, and kernel rank in one. Under a change of drive band the ordering reverses in 10 of 10 seeds in both, and the band at which the crossover occurs moves with the delay horizon. The systems compared differ only in their time constant, so this is a demonstration of an effect the field already predicts, not a discovery; the contribution is the measurement and the reporting condition that follows. The note recommends that any published ranking state its observation band, its delay horizon, and which axes under which protocol. Includes zharnikov-2026bq-band-conditionality.yaml (Paper Spec v0.1.0) – a machine-readable specification of the paper's claims, assumptions, and dependencies. The paper's full machine-first bundle (the SPINE claim/dependency graph and the ONTOLOGY term module) lives in the public repository; see https://github.com/spectralbranding/paper-spec for the standard. This PDF is generated programmatically from that machine-first source under a research-as-repository model.