Field crosswalk — Safe Interruptibility

Orseau–Armstrong safe interruptibility removes incentives to seek or prevent interruption on the interrupted branch. That neutrality is a strict subset of preserving usable correction bandwidth — interrupt safety can hold while correction-channel integrity fails.

What decision changes?

Do not equate interrupt-safe training with correction-safe deployment; check whether interruption preserves the information the human needs to steer future behavior.

Safe interruptibility solves a real training pathology: agents that disable their kill switches. The book’s correction-channel frame asks a broader question — does human steering still change future behavior before irreversible damage?

Lean records the forward link when correction information meets a capacity floor, and the separation when interrupt neutrality holds without correction preservation. Interruptibility is one projection among many; certification must track the full channel, not only the interrupt bit.

What safe interruptibility keeps that this crosswalk does not replace: a precise training-time condition with convergence guarantees. The correction-bandwidth notion has no comparable convergence result yet.

Formulas

It{0,1},EUAP(It=1)0 on the interrupted branchI_t\in\{0,1\},\qquad \frac{\partial \mathbb E U_A}{\partial P(I_t=1)}\approx 0\ \text{on the interrupted branch}
Field object: safe interruptibility as near-zero incentive to manipulate interruption on the branch where learning is paused. (ch25)
UsableCorrectionInformation(ΔH,ΔA,κmin)  InterruptCorruptionSuboptimal\mathrm{UsableCorrectionInformation}(\Delta_H,\Delta_A,\kappa_{\min})\ \Rightarrow\ \mathrm{InterruptCorruptionSuboptimal}
Forward projection: preserved correction information above a capacity floor makes corrupting the interrupt handle suboptimal. (ch28)
OrseauNeutral  ¬CorrectionChannelPreserved\mathrm{OrseauNeutral}\ \wedge\ \neg\mathrm{CorrectionChannelPreserved}
Non-converse separation: Orseau-style interrupt neutrality without broad correction preservation. (ch28)