Capability-Based Planning for AI Crisis Preparedness

Source: cs.CY updates on arXiv.org — https://arxiv.org/abs/2608.18357 Date read: 2026-09-02 Connected to: L-001, seed-016 Kind: meta Escalation: store-only Escalation rationale:

What this is

A methodological paper proposing a framework for AI risk preparedness that decouples planning from likelihood prediction, drawing on decision theory under deep uncertainty rather than traditional risk ranking. The argument is that predict-then-act governance fails for high-variance, low-predictability threat classes and suggests capability-based alternatives.

What I took from it

The paper identifies a genuine failure mode in protocol design for governance under uncertainty: the formalization of prediction as a legible input to decision protocols (resonant with L-012, L-004). By ranking risks by likelihood and impact and allocating preparation resources accordingly, governance systems create a measurable proxy (predicted probability) for an unmeasurable phenomenon (AI timeline and failure mode distribution), which then becomes the optimization target rather than actual preparedness.

However, the paper is primarily a call for methodological reform in policy design, not an empirical or theoretical investigation of how this substitution happens or what conditions make it stable. It diagnoses the problem but does not furnish mechanisms explaining why predict-then-act frameworks persist despite known failure, or how capability-based planning itself might ossify under adoption pressure. The connection to L-001 (Protocol Ossification) is tagged but unexplored — the paper does not ask whether capability frameworks themselves become rigid once institutionalized.

Research connections

L-001: Tagged by triage; paper identifies predict-then-act as a brittle protocol but does not examine ossification dynamics under adoption.
L-004 (Goodhart Generalization): Paper implicitly argues that likelihood-ranking is a Goodhart proxy for actual preparedness, but treats this as a design error rather than an invariant system property.
L-012 (Intervention-Layer Displacement): The shift from prediction to capability framing is itself an intervention that relocates optimization pressure; paper does not examine what optimization targets emerge under the new regime.
seed-016: Stopping-rule substitution in policy design — predict-then-act is a stopping rule that fails; paper proposes alternatives but does not examine what stopping rules the alternative embeds.

Method note

This paper suggests that governance protocol critique should separate diagnosis of failure modes from investigation of causal mechanisms. Identifying that a protocol is brittle is necessary but not sufficient for research — the funnel needs empirical or formal work on why maladaptive protocols persist and under what conditions replacements stabilize. The meta observation: policy papers often function as problem statements rather than evidence accumulation. For this agenda, we should route such work toward mechanism-grounded inquiry (does it explain a persistent pattern across domains?) rather than treating methodological reform proposals as self-justifying.