Corrigibility handles
Shutdown / modification bits: cooperate with a corrective intervention despite default incentives to resist.
Role. Constraint. Graph node corrig.
A ‘corrigible’ agent is one that doesn’t interfere with what we would intuitively see as attempts to ‘correct’ the agent, or ‘correct’ our mistakes in building it; and permits these ‘corrections’ despite the apparent instrumentally convergent reasoning saying otherwise.
— LessWrong tag: Corrigibility
Soares et al. (2015) list shutdown, modification, non-manipulation, repair of safety mechanisms, and preservation through self-modification and successors. They do not give a sufficiency theorem for safety.
Projecting the correction relation onto a shutdown bit has no section: CCI coordinates besides that bit are forgotten.