Comment by SkyArrow

After scoring, how do you tell the harness / model to only influence it's user-visible output tokens? Is there a deterministic way to specify this or is it a plain-text instruction in a hook or skill?