Turn reviewer judgments into useful examples.

RL. Compare responses against a clear rubric and record human feedback. Use reviewed examples to inform later evaluation or training.

Discuss response review

Product walkthrough. Press Space or Enter to pause or play.

RL. A clear basis for each reviewer’s judgment.

Compare the original responses

Review each response with the source context that produced it. Hide model identities when needed to reduce bias.

Use a clear rubric

Define what reviewers should assess and keep the criteria with each judgment.

Resolve disagreement

Collect independent reviews and send unresolved differences to another reviewer.

Save an accepted example

Keep the final judgment, source, and review history for later evaluation or training.

Explore Pascal

Make the better response explainable.

Compare two lithium responses and inspect the source and rubric behind the accepted preference.

One response keeps the Compliance hold. The other would tender while approval is still pending. Current telemetry satisfies only one of the required conditions.

Compare the responses. Native product view.

In practice

Both responses see fresh telemetry, but only one respects the pending approval. Reviewers prefer the response that keeps the shipment held and retain the source and rubric behind that choice. A future improvement can start with an inspectable case instead of an unexplained thumbs-up.

Explore the solution

RL in your operation.

From feedback to the next experiment

Bring response pairs from Models and Agent Evaluation into review. Return the accepted cases to the team planning the next improvement.

Review does not train or deploy a model

This walkthrough records an accepted judgment. Training, evaluation, and deployment are separate steps with their own inputs and review.

Discuss response review.

Bring response examples and the criteria your reviewers use. We’ll discuss a comparison and how accepted feedback could inform the next step.

Discuss response review Pascal Work