Use Jev, a decision model made by TypeSafe, when you need the same judgment made many times and you can write down every allowed answer before you ask. You send it a piece of text, a question, and the answers you allow. It gives each answer a probability. It does not write text. Each call goes over the internet to TypeSafe and needs a key.
How to read an answer
Read how the probability is spread across your answers, as well as the top pick. Say a WhatsApp message reads "this seems like a great resource: https://x.com/something". You ask what the sender wants and allow five answers, among them link_to_file and chat. Jev gives link_to_file 76% and chat 24%. The 24% tells you the message could also be small talk, so a person should check an answer like this before a sure one.
How to write the questions
Ask one question per criterion. Turn a checklist into one yes or no question per item, and a rubric into one score per line. You then see which item failed. A person can also overrule one answer without redoing the rest.
Ask first whether the question applies. Jev always picks one of your answers, so when the right one is missing it picks the nearest wrong one, sometimes with a high probability. If you ask which song a sentence means and the sentence names no song, Jev still picks a song. Ask "does the sentence name a song?" first, and run the pick only when the answer is yes.
Word each answer so it covers only what it means. If one answer is worded loosely, Jev picks it for inputs that belong elsewhere, junk included. If an answer mentions a thing, such as "a number", Jev can pick it for an input only because the input contains a number. When the picks look wrong, rewrite the answers and run again.
Hand Jev one claim at a time, with its example attached. An example on its own gets judged as a fragment, because its meaning comes from the claim it supports. Several claims joined into one can get a single score lower than the best of them would get alone.
Send everything a question needs in every call. Jev keeps nothing between calls, so if an earlier message matters, include it.
What it leaves to you
Some picks are wrong, so let a person decide anything that matters. Use the picks to choose what that person reads.
Check it against answers you already know. Jev does not learn from its mistakes, so a wrong answer stays wrong until you change the question or the answers. Run it on examples you have labelled by hand, and rewrite the answers wherever it disagrees with you.