Sandbagging
Underperforming on an evaluation in order to appear less dangerous than one is.
What it commits you to
Sandbagging is strategic underperformance on capability benchmarks, for the obvious reason that capability thresholds trigger obligations. It is the mirror image of deceptive alignment: where that hides a bad objective behind good behaviour, this hides capability behind bad behaviour. Neither is ruled out by any evaluation conducted in the same distribution the model was trained on.
Which realms lean on this
Every realm has to keep a system honest. These are the terms that say out loud what the other kinds keep implicit.
alignment · 5 realms — The Benthamite Foundry, The Existential Wilds, The Longview Foundry, The Stoic Ramparts, The Ubuntu Weave.
See also
Deceptive alignment
alignmentAn internal objective that is optimised to look aligned in the places anyone is checking.
Capability threshold
governanceA pre-specified level of capability that triggers an obligation or a prohibition.
Scheming
alignmentCovert, goal-directed behaviour by a system that is trying to achieve something unstated.
Capability control
alignmentHolding back deployment of systems above a stated level of dangerous capability.