Judgment needs a standard.

Ask two skilled reviewers to assess an answer and they may notice different things. One may focus on factual accuracy; another on whether the answer resolves the request. Their disagreement can reveal something useful about the task, provided the review process gives them a shared language.

A rubric begins by separating requirements. Is a claim supported? Were the relevant constraints respected? Does the result give its intended reader enough information to act? These are different questions, and collapsing them into a single impression can hide the reason one output is preferred.

The standard also needs examples at its boundaries. A clearly unacceptable result is easy to reject. The more instructive cases are plausible answers with one consequential omission, or correct answers that rely on an assumption the task never allowed. Discussing such cases can sharpen the criteria before they are applied widely.

Automated checks have a role when a requirement is observable and well specified. Human review remains useful where context changes the meaning of an otherwise reasonable response. The aim is to make the division of responsibility explicit.

At Silo Data, we see evaluation design as part of data design. The criteria used to assess an example influence what gets created, revised, and retained. A thoughtful standard gives those decisions a traceable basis and makes the next iteration easier to direct.