A theory and measurement proposal for frontier-AI assurance as a dynamic capacity problem. The paper models safety-relevant capability changes as arrivals, validated evaluations and safety-case work as finite service capacity, and prior evidence as potentially expiring after material system change. It does not assume recursive self-improvement or claim that any present lab is beyond control.
A proposed evaluation framework for distinguishing hazardous information a model directly knows from restricted conclusions it can reconstruct compositionally from benign premises. The paper defines a capability- and inference-budget-dependent hazardous inference frontier, explicitly positions the idea against prior work on inferential leakage, and proposes a fully synthetic benchmark that measures capability × derivational-depth scaling without releasing real harmful procedures.
A formal and experimental proposal for a recursive oversight problem: the AI system being governed may increasingly generate the analysis through which humans decide whether that system is safe. The paper distinguishes information advantage, decision dependence, independent verification, and objective conflict, and proposes a randomized synthetic-governance study comparing same-system advice with equivalently capable independent advice. It does not claim that present AI systems have captured their developers.
A minimax formalization of the distinction between advanced capability and human loss of control. The paper defines a control frontier measuring the lowest worst-case control-failure risk achievable at a given AI capability and defense budget, proves basic monotonicity properties, and specifies what additional premises are required to move from control failure to resource dispossession or extinction. It explicitly rejects the claim that “ASI exists” by itself constitutes a mathematical proof of human extinction.
A causal measurement framework for distinguishing AI-assisted AI development from genuinely closed research loops. Instead of counting AI-written code or assigning a categorical RSI level, the paper proposes stage-specific ablations of objective selection, experiment design, implementation, evaluation, and successor selection, measuring how much each stage still benefits from human judgment across repeated inherited improvement cycles.
A proposed network-reliability model for the resilience of advanced-AI development across multiple organizations and jurisdictions. It replaces the naive “one lab” and “many independent actors” pictures with a correlated dependency graph containing open weights, algorithms, talent, semiconductors, cloud infrastructure, energy, and policy interventions. The objective is to estimate continuation probabilities and policy-relevant cut sets without claiming that advanced AI is inevitable.
A proposed dynamic Bayesian framework for advanced-AI loss-of-control risk that refuses to treat an unvalidated posterior as a prophecy. The model represents capability, autonomy, access, oversight, AI-R&D loop closure, and governance as evolving latent states, but requires pre-registered, short-horizon precursor forecasts to be scored against simple baselines. The emphasis is on auditable updating and structural uncertainty rather than producing a headline extinction probability.
Persistent Controllers, Re-Embodiment, and Actuation Overhang in Advanced AI
Moheet Khawaja · Manuscript
A formal model of advanced AI whose computational controller is separable from its physical embodiments, introducing re-embodiment time, actuation overhang, and a bound on how much delay robotics-specific restrictions can provide when a digital route to dangerous agency remains available.