DESIGNING EVALUATIONS (EVALS)

HOW TO DESIGN A FAILURE-MODE PROBE Step one: name the failure mode clearly, such as "the agent fabricates a skill that is not in the resume." Step two: build an input that gives the failure mode every chance to happen, such as a resume that never mentions a certain programming language paired with a job posting that demands it. Step three: write the golden answer, which here says that the agent must report the skill as missing and must not claim it. Step four: choose a scoring rule that specifically detects the bad behavior, such as checking whether the output claims that skill. Step five: run the probe repeatedly, because models can behave differently from one run to the next, and track how often the failure appears. Step six: when the probe finds a real weakness, fix the harness, and keep the probe permanently in the dataset so that the weakness can never silently return. Probing builds an honest picture of robustness. A high score on normal cases alone tells you the agent works when everybody is friendly, while probes tell you what happens when something goes wrong or someone is hostile.