Judge Role

The distinction between asking a model to generate candidates, to judge a relation, and to arbitrate which belief survives — the last of which it must never do.

Why it matters in a memory layer

Judging a retrieval answer is comparatively safe: the judge scores against documents that still exist, a human can read both, and a wrong call produces a wrong number. A judge in a memory system decides what is stored — its output is state, not a score, and a wrong call retires a true belief permanently with no reason anyone can inspect.

So the split this course made two levels earlier is load-bearing: the model says these two disagree, and rules say this one wins. Detection is a language question; arbitration is a policy whose output changes what is believed and therefore has to be explainable. A model there produces a store whose contents depend on sampling, and "why do you think that?" has no answer.

Generation is judged differently again. This system's two generation sites are unbounded and unscored, and that is acceptable because nothing downstream trusts them — four later stages exist precisely because extraction over-produces.

Connections

Taught in: LLM as Judge, and Its Failure Modes

Do not confuse with: Memory Operations