Alignment as a Measurement Problem

The first alignment question is not whether a system is good, but where the effective optimizer actually is.

What decision changes?

Before evaluating goals, corrigibility, or deception, ask which bounded subset of the world — in the operational, not metaphysical, sense — best predicts and controls the outcomes at stake.

Many alignment arguments begin too late. They ask whether a model’s objective is safe, whether a policy is corrigible, or whether a trained network contains a mesa-optimizer — all of which assume we already know what object is being aligned. Often we do not.

The book reframes the opening question as operational: which subset of the world, modeled as a bounded process with internal state, interface channels, and control capacity, best predicts the future changes that matter? That is a measurement question before it is an evaluation question.

This reframing carries through the rest of the book. Boundary discovery, capability measurement, and correction-channel integrity are all downstream of first locating the right object — a step skipped by approaches that start from “is this policy’s behavior acceptable?”