Four of six AI evals could never fail

Alwayspass found that four of six test cases in its broken Promptfoo fixture could never fail, leaving only 33% of the suite able to go red. The new static-analysis tool flags model-graded assertions without thresholds, tests with no assertions, vacuous checks, and expected answers copied into prompts without running a model or spending tokens. Teams applying the whole-system evaluation discipline demanded by agent harnesses should put this kind of falsifiability check before paid evals in CI: a green gate is evidence only when some output can close it.