Evaluation before capability

Without a stable evaluation, an improved result is indistinguishable from luck. Progress becomes anecdote.

Building the evaluation first forces a definition of the work. What counts as finished? What does failure look like? Which failures end the run? Those answers are the specification. They outlast the model.

Evaluation grounds decisions about authority. What has the agent demonstrated? Under which conditions? What happens when it leaves them?

Evaluation starts with the work and stays in the loop. As capability grows, the test must keep pace.