02
Evals
Every change, scored.
Replay your test cases against any version and see the pass rate, the regressions and the exact cases that moved.
THE PROBLEM
A prompt change looks fine in a demo and breaks an edge case nobody remembered. Without a repeatable test, quality is whoever tried it last.
HOW TRACEWELL HANDLES IT
Evals are plain code or plain-language rubrics. Run them locally, in CI or on every deploy, and diff the scores against main.
01
How it works
Three steps from nothing to useful.
01
Build a dataset
Add real runs as cases, redacted by your rules.
02
Write a check
A function that returns a score from zero to one, or a rubric.
03
Gate the merge
Fail the pull request when the pass rate drops.
+1.6 pts
pass rate gained against main in v2.4
0
new failures reached production last quarter
Try it on your own agent.
Free for 10,000 runs a month. No card, and the first trace takes five minutes.