Pointing Problem

The field’s pointing problem is an umbrella, not one technical task. It is the family of questions about whether a system is aimed at the right human-relevant target, and whether that aim survives construction and later change. This project splits it into three senses that fail independently.

Identification asks what the target is: which values, which bearers those values apply to, and which correction process is legitimate. The precise bet is value/bundle identifiability (MB2; field noun Value Learning). Bearer maps (MB3) sit next to that slice: who the values apply to is adjacent to identification, not the same crux. Wentworth’s Pointers Problem is the identification-heavy slice: human values as a function of latent variables inside a human world-model.

Realization asks how to build a system that tracks that target. This project discusses relevant mechanisms but does not claim a general construction method. That remains an open interface, not a bridge assumption. Hubinger’s outer-alignment “pointer” is a third-party label for a mix of identification and construction; it is not identical to any one of these three senses.

Preservation asks how the system keeps tracking: whether the correction channel , transport  layers, and later certification still bite after capability growth, ontology shift, successors, and selection. The precise bets are correction-channel integrity (MB4, MB4a) plus layered certification.

A hospital can write “patient welfare” on its charter (identification), fail to hire or train anyone who can act on it (realization), or keep the charter while quietly optimizing billing codes (preservation). Those are three different failures, even though people will call all three “not pointing at patients.”