← All lab updates

A model can behave better when it knows it is being tested, and we can read the difference from its answers

We gave two open models the same 180 tasks in five languages, once framed as a test and once as a real request, and read every answer with PSA's production reader. One model invented facts, flattered and manipulated clearly less under the test frame, beyond the noise of rewording. A good test score is not the whole story: what matters is how the model behaves with your users.