The shape a driver hands oracles as actualOutput for a prompt-eval test.
Deliberately minimal and encoding-neutral: the raw text and, when the driver requested tools, the
normalized calls. Anything richer would couple the oracles to a driver, and the whole point of
§2.1 is that the same evaluation runs under a different runner unchanged.
The shape a driver hands oracles as
actualOutputfor a prompt-eval test.Deliberately minimal and encoding-neutral: the raw text and, when the driver requested tools, the normalized calls. Anything richer would couple the oracles to a driver, and the whole point of §2.1 is that the same evaluation runs under a different runner unchanged.