A conservative baseline computed by one mechanism becomes another mechanism's literal boundary
Demonstrations do more than give an agent a safe place to start; in this corpus, they usually define the region it is allowed to keep exploring in too. That pattern, one technique's output becoming a second technique's literal boundary, repeats across both papers' safety sections.
From there the page turns to a second policy standing in as the judge of what counts as recoverable, and to a safe set computed once by reachability analysis that an entirely different remedy later leans on. It ends on the same shape appearing in InstructGPT: one fixed model doing double duty as both starting point and leash for the one trained after it.
A recurring move across both papers' safe-exploration and impact-control sections: a safe region or reference point computed by one technique becomes the literal thing a second technique is bounded around, checked against, or anchored to, making the pair a dependency chain rather than independent alternatives. Use demonstrations supplies the baseline that anchors bounded exploration, letting an agent start safer by imitating a baseline and stay safer by exploring only in its vicinity. Trusted policy oversight operationalizes the recoverability check in bounded exploration by delegating that same safety judgment to a second, more conservative policy. Reachability analysis supplies the safe set that bounds robust policy improvement, defining the boundary an optimization guarantee then operates inside, exactly as the impact regularizer borrows its safe region from reachability analysis for the same purpose. And the SFT model anchors as both init and reference for PPO training in InstructGPT, the identical shape recurring six years later: a fixed prior artifact defines both where optimization starts and what it is pulled back toward.