Definition
Pointwise evaluation assigns an output its own score against stated criteria. The evaluator uses the context needed for the task, such as the original request, source material, or an expected answer. The score reflects how that output meets the criteria without directly comparing it with a competing output.
A person, a model judge, or a deterministic check can assign the score. The result might be pass or fail, a number on a defined scale, or separate scores for different criteria. The evaluator may use a reference answer, but does not need one. Each output is assessed on its own, even if the team later compares scores across prompts or models.
Simple example
A support assistant answers a question about refunds after 30 days. The policy says refunds are available only within 14 days. A reviewer scores factual accuracy from zero to two, with zero for an answer that contradicts the policy. The assistant replies, “Yes, you can get a full refund” and receives zero. The reviewer needs no competing answer to reach that score.
Why it matters
Pointwise scores let a team assess a deployed system without producing a second answer for every test case. The team can track the share of answers that meet a minimum accuracy standard and inspect the failures.
When testing a change, score both versions on the same cases with the same rubric. You can compare their results afterward, while each individual judgment remains pointwise.
One important nuance
A number can look more precise than the judgment behind it. If “mostly correct” has no clear definition, two reviewers may score the same answer differently. A model judge can also shift its scores when its instructions or model version change. Define what each score means, check borderline cases against human review, and keep the rubric and judge version with the results.