Skip to content
Tracks @XLab

The diffuse-to-concentrated frontier

Move the deployer's three mitigations and watch which strategies survive and where the schemer's best response migrates.

The diffuse-to-concentrated frontier

Move the deployer's three mitigations and watch which strategies survive and where the schemer's best response migrates.

The schemer's Pareto frontier, colored by the strongest mitigation

concentrated: few actions, each blatantdiffuse: many actions, each innocuousblatantsubtle1101001,00010,000malign actions needed for catastrophe

Probability each frontier strategy survives all three mitigations

100%0concentrateddiffuse
incriminationtraining on flagged actionsuntargeted training

Schemer's best response

1 action

Its survival probability

96%

Binding mitigation there

none of these

When every lever is turned up, the surviving strategies crowd into the far concentrated corner — so few actions that random review never sees one. That region is not covered by these mitigations at all: it is the territory of the blocking-and-auditing techniques from the original control paper, which is why deployers need both toolkits at once.

Embed this demo

Use the chrome-less embed view inside an iframe:

<iframe src="/demos/sabotage-frontier/embed" width="100%" height="360" style="border:0"></iframe>