Measured Lab
Measured Lab is open. Its three tabs run in your browser.
- What it is
- A web app where you check AI work against rules you can read, so you see why something passes or fails.
- Evaluate
- Pick one of 20 test topics, read its rules and sources, and check two drafts side by side on the same rules. Change a draft and check it again.
- Observe and Secure
- Observe reads an agent's run step by step, from a recorded example or a trace you paste. Secure screens your own text for instructions hidden in it, and checks what an agent would send out.
- Lab result links
- After a check, you can make a link that records exactly what was run. Anyone who opens it can run it again in their own browser and see whether the result matches. The result lives in the link; nothing is stored on a server.
- How it checks
- The checks are code from measured-evals, the open-source evaluation package of Proofline. They run in your browser: no model, no key, and your text is not sent anywhere.
- Not done yet
- Checks that need a model or a person are shown as not run. There is no Pro plan and no price.
Open Measured Lab (opens another site)
See alsoBuild log