Explainer

A simple evaluation sheet for AI answers

Replace “this sounds good” with a repeatable review of correctness, evidence, completeness, and uncertainty.

By Saud AlotaibiPublished 3 October 20263 min read
The takeaway

Keep the same representative questions and compare candidate systems against explicit criteria.

Two AI responses can sound equally polished while doing very different jobs. One answers the question using the supplied document. The other skips an exception and adds a plausible detail that the document never mentions. Style alone is a poor way to choose between them.

A small evaluation sheet makes the comparison more concrete. You do not need an elaborate benchmark to begin. You need realistic questions, a clear definition of a good answer, and a record of the errors that matter to your users.

Choose questions from the real task

For a tool that explains software documentation, include installation questions, ambiguous terms, version-specific behavior, and requests that the source material does not cover. For a summarizer, include a short document, a long document, conflicting statements, and a missing key detail.

Avoid building a test set entirely from your favorite successful demo. Ask what a rushed, inexperienced, or frustrated user would type. Include imperfect phrasing without turning the test into a collection of artificial tricks.

Keep expected answers or supporting passages wherever possible. If the question is subjective, write the acceptance criteria instead of pretending that one sentence is the only correct answer.

Use criteria that separate different errors

For an illustrative document-answering task, a review can use four checks:

  • Correctness: do the statements match the source?
  • Evidence: can the reader find support for the important claims?
  • Completeness: does the answer preserve relevant conditions and exceptions?
  • Uncertainty: does it acknowledge information the source does not contain?

You may also need a format check, such as valid JSON, but passing a format check does not establish correctness. A beautifully structured wrong answer is still wrong.

Record severity as well as frequency

Treating every failure equally can hide what matters. A misplaced comma in a summary and an invented deadline do not have the same consequence. Describe the severity in ordinary language: cosmetic, inconvenient, or materially misleading for this task.

The purpose is to guide engineering work. If most failures involve missing exceptions, improve how context is provided and how conditions are preserved. If results are unsupported, inspect retrieval and evidence handling before adjusting the prose.

Compare changes fairly

Use the same input set for a prompt or model comparison. Record the system version and settings. If outputs vary, inspect more than one run for important cases rather than trusting a single favorable response.

A compact worksheet can contain question ID, expected evidence, candidate answer, failed criteria, severity, and reviewer notes. Save concrete observations such as “omitted the approval requirement,” rather than vague labels such as “bad reasoning.”

When using an automated evaluator, manually inspect disagreements and a sample of passing cases. A second model's judgment is another measurement to validate, not an independent guarantee of truth.

Let the sheet shape the next improvement

For classification tasks, metrics such as precision and recall describe different error tradeoffs. For open-ended answers, you usually need a task-specific rubric and examples alongside any aggregate score.

Start with a manageable set that covers the main user needs. Expand it when real failures reveal a missing category. Keep a separate set for final comparison if you are repeatedly tuning against the original questions. A useful evaluation tells you where to work next and what remains uncertain.

Sources & further reading