Start with three kinds of criteria

Structural criteria can be checked mechanically: a required section exists, a minimum number of accessible sources is included, or the output follows a schema.

Semantic criteria require judgment about meaning: whether a brief covers the important subtopics, whether a conclusion is supported, or whether a summary preserves the source’s argument. These checks need an explicit evaluator and evidence.

Human criteria reflect authority: whether the evidence direction is appropriate for the decision, whether a risk is acceptable, or whether the final result should be delivered.

Make each criterion observable

A useful criterion points to something the reviewer can inspect. Instead of writing “the research is comprehensive,” write “the brief covers pricing, distribution, and customer demand, with at least two sources per subtopic.”

Instead of writing “the result is accurate,” define what support is required. For example: “Every material market claim maps to a named source and a quoted or summarized passage.”

Do not hide human judgment

Some decisions cannot be reduced to a model score. If a person must own the risk or direction, encode that as a human checkpoint rather than asking an evaluator to imitate authority.

A human criterion might be: “The user confirms that the evidence supports the intended market thesis before final drafting.” That creates a clear pause with a clear owner.

Use a small, testable template

The goal is not to write a perfect legal document. It is to make the difference between “done” and “acceptable” inspectable before execution starts.

  • Result: What must exist at the end?
  • Structure: What must be present or follow a schema?
  • Evidence: What must support each material claim?
  • Boundary: What may not be used or changed?
  • Human decision: What requires explicit approval?