Find the LLM calls that are really decisions
Route this ticket, flag this comment, score this lead: plenty of LLM calls are decisions in disguise. Paste one, and Jev says whether a judge can take it over, and which.
An LLM call from your code
Paste a prompt and check it. You'll see whether a judge can take the call over, and which one.
Three signs a prompt is a decision
A builder who scanned over 1,000 repositories found calls like these everywhere. The check above looks for the same signs.
The reply gets parsed
Your code strips the reply, lowercases it and branches on it. Nobody reads the words; they only carry a value.
The answers are a short list
Yes or no, one of five teams, a score from 1 to 10. When every valid answer is known in advance, it's a decision.
You fight the format
"Reply with only the category name" and a retry when it doesn't. A judge can't answer outside its options, and says how sure it is.
Swap one call this week
Every ready-made judge runs from one API call. 500 judgments a month free, no credit card.