Example context. Preview only. Nothing is submitted.
When to sendReviewers can read the same source text and both summaries, and understand the criteria for a useful summary.
- 01
Read and compare
Keep the source and both outputs available while the reviewer decides. A page break that hides the text would turn the task into a memory exercise.
See the questions
- Which summary does the better job?
- Distinguish a preferred summary from a tie or an inability to judge.
- What led to your choice?
- Capture the evidence behind the choice without making a comment mandatory.
Using the answers
Keep each judgment tied to the source, both outputs, and the reviewer assignment. Report ties and unable-to-judge answers separately. Compare balanced, blinded assignments across a defined evaluation set; do not infer a model win rate from one static pair.
When to choose another form
Treating one preference as proof of factual accuracy, general model quality, or statistical significance.
Before you publish
This shared form reviews one fixed pair. For a model comparison, assign multiple cases, vary A/B order, hide model identities, and store the mapping outside respondent-visible copy. The template does not randomize assignments.
Train reviewers on the task, check agreement on a sample, and inspect explanations. Feedback collection does not automatically train a model.