Skip to content

The control graph

Build the high-stakes control graph region by region — high-stakes, low-stakes, the trivially-solved and probably-hopeless corners — then see why incrimination and the number of actions are inversely correlated.

High-stakesLow-stakesTrivially solvedwith random auditingProbablyhopeless?fewer actions →bigger, more incriminating actionsHow incriminating is each action?Number of actions to cause catastrophe

1. The axes

The control graph places a deployment scenario on two axes: how incriminating each of the model's actions is, and how many actions the model needs to cause a catastrophe.