Evals
Checks test your code. Evals test the harness. Change the words an agent reads and nothing goes red, so this is what tells you whether the change made it better or worse.
Run it with /hyperpower:eval. It prints SHIP or HOLD.
Read them in order
- What an eval isA test suite for prompts
- What a case looks likeA prompt plus what a good answer must contain
- How to run itName the prompt under test, then read the verdict
- Who writes the casesYou do, from work you have read
- How it decidesSix weighted dimensions, three conditions, SHIP or HOLD
- The reportOne page showing every prompt and its baseline