See where people give your AI a

Ratings, answer quality and comments over time. Every update is marked, so you can see what changed.

In short

Over time

Weekly average · all topics

Where people are unhappy

Share of 👎 by topic. Tap one to focus on it.

Why they're unhappy

Number of 👎 by reason. Darker means more.

What each update changed

👍 rate the week before vs. the week after

Read what people said.

Every 👎 is a clue. Filter by topic, reason or version to find the pattern.

Test before you ship.

AI can change in surprising ways. These tricky questions catch problems like made-up facts or leaked personal info before your users do.

–passed

Ready to test

Checks the new version against the one people use today.

Test your own AI

Add your own tricky questions to evals/cases.js. For each one, write what a good answer must (or must not) say.

Run them against your model:

PRODHEALTH_API_KEY=… node evals/run.mjs \
  --endpoint https://api.example.com/v1/chat/completions \
  --model my-new-model \
  --baseline evals/results/baseline.json \
  --out evals/results/candidate.json

If something important breaks, the command fails, so it can stop a release automatically. Open More → Load new version results here to look through the answers.