Feasibility of a Joint Risk-Ceiling and Development-Deadline Guarantee

Determine whether a single regulatory policy can guarantee both the specified hazard ceiling and development deadline when the admissible lower bound on safety-research productivity lies between the necessary productivity bound and the learn-then-replay productivity bound.

Background

The paper studies regulatory commitments that simultaneously impose a ceiling on AI hazard and a maximum allowable delay relative to a full-speed development path when safety-research productivity is uncertain. It derives a necessary productivity bound that any compliant policy must satisfy and constructs a learn-then-replay policy with its own sufficient productivity bound.

The authors explain that the interval between these two bounds represents unresolved feasibility: productivity above the replay bound supports the constructed guarantee, while productivity below the necessary bound makes any such guarantee impossible. When available evidence places the minimum admissible productivity strictly between the two bounds, the paper does not determine whether some policy outside the analyzed replay class can satisfy both commitments.

References

The binding input is the productivity bound. Evidence that rules out productivity below the replay bound turns an impossible universal promise into a feasible conditional one; evidence that stops between the two bounds leaves the question open.

— Risk Ceilings and Development Deadlines: Pacing AI under Uncertain Safety Productivity  (2610.11093 - Gan, 8 Oct 2026) in Section VI, “Policy Commitments and Their Scope”