# Pascal Agent Evaluation

Source: https://www.trypascal.io/platform/evaluation

Know whether the next version is better.

Compare agent releases, models, and instructions on the same cases. Inspect regressions, bring in human judgment, and carry the evidence into your release decision.

## Capabilities

### Build useful datasets

Author or import cases and curate captured failures. Keep reference answers, source lineage, and dataset versions available for review.

### Define what good means

Combine deterministic checks, Python scorers, model judges, and classifiers. Publish exact evaluator versions and test them on examples.

### Compare variants in a playground

Run the same cases across models, instructions, or exact Agent Releases. Inspect outputs and scores, then rerun the cases that need attention.

### Look beyond an average

Compare paired outputs and relevant slices. Review uncertainty, regressions, and missing results before accepting a candidate.

### Bring in human judgment

Use independent reviews and adjudication to resolve disagreements. Compare judge results with accepted human labels and inspect calibration coverage.

### Turn failures into regression cases

Explore behavior distributions and failure groups. Review examples, retain accepted expectations, and publish cases for the next evaluation.

## Connected context

Start with captured evidence in Observability. Bring in independent human review through Lab. Carry the evaluation into the next agent draft and release decision.

- Versioned datasets and frozen evaluation samples
- Evaluator versions and exact agent variants
- Runs, paired comparisons, and human review

## Controls and permissions

A comparison retains its exact cases, evaluators, and targets. Missing results remain unscored, and a favorable example does not authorize a release.

## Example

The support team compares releases 3 and 4 against refund cases. The candidate requests approval on the disputed case, but another planned result is missing. The team keeps the comparison inconclusive and collects more evidence before making a release decision.

## Workflow

### Find the failure

- Locate the captured refund promise and the calls around it.

### Inspect the evidence

- Read what the application reported and distinguish the response from a confirmed refund.

### Prepare a regression case

- Keep the original request and supply a reviewed expectation: check the policy and request approval.

### Compare two releases

- Inspect the same case on the baseline and candidate, including their different responses.

### Review the decision

- One candidate result is missing. The comparison remains inconclusive and calls for more evidence.

## Next steps

Choose an agent, the failures that matter, and the criteria your team would trust to judge an improvement.

- [Contact sales](https://www.trypascal.io/contact)
- [Explore Agent Observability](https://www.trypascal.io/platform/observability)