The review partitions the field into five research communities that rarely cite one another. Each pillar has its own search string, its own paper table, and its own list of what it leaves open for PreCog.
How hazards arise from the interaction of many agents, and how multi-agent state can be represented tractably enough to evaluate safety over it.
PILLAR 2 · 6 PAPERSNon-generative world models that predict in latent space, and graph or hyperbolic representations suited to hierarchical enterprise structure.
PILLAR 3 · 9 PAPERSWhen a model should abstain or hand off to a human or safeguard, from the reject option through consistent surrogate losses to sequential and multi-expert deferral.
PILLAR 4 · 0 PAPERSRunning learned models on edge silicon under hard latency budgets, and the effect of wireless tail latency on time-critical decisions.
PILLAR 5 · 6 PAPERSWhat a human supervisor needs to take over from automation, how automation erodes that capacity, and who carries responsibility when it fails.