Kolaboro / Engineering / Evaluation
Daily AI evaluation
Every morning the same code that serves buyers reads the same fictional test pages, and each answer is checked against fixed expectations. If the model, the prompt or the page parser changes and something breaks, it shows here, and we get an email.
Every case
| Case | Kind | Result | Checks |
|---|---|---|---|
| Loading… | |||
History
| Day | Cases | Checks | Model | Code |
|---|---|---|---|---|
| Loading… | ||||
How to read it
Each case is one fictional page in /open/evals/fixtures/ with checks listed in cases.json: a field of the answer and what it must be. A case passes only when every check passes. The run records the fingerprint (SHA-256) of the prompt and of all fixtures, so a change in either is visible in the history. Raw data: /api/open/evals, signed like every open answer.
More from Engineering: AI that can't be talked into a verdict · Daily AI evaluation · The public log · Signed site photos · Reliability