Most LLM classification work happens by asking a chat model a question and parsing free text out of the response. Ask “is this urgent” and you get a paragraph, then you’re regex-parsing a sentence to get a boolean out of it. It works until the model phrases something differently or wraps the answer in a caveat.
TypeSafe’s evaluation API (the model is called Jev) skips that step. You send it a state (the text you want evaluated) and a map of typed questions, and it returns typed, probability-scored answers, no parsing required. We tested it against three real quote-to-cash scenarios, an order that deviates from standard payment terms, a disputed invoice, and a collections thread with mixed signals, and pulled the actual API responses rather than hypothetical ones.
Get RevExOS’s stories in your inbox
Join Medium for free to get updates from this writer.
The most useful result wasn’t the confident one. Feeding it an ambiguous collections thread returned a score sitting exactly between “medium” and “high” urgency, with confidence at only 0.25, which is the model correctly refusing to be more certain than the evidence supports. That’s the behavior you want before anything auto-escalates a customer account.
We also cover where it fails (arithmetic and date comparisons, by design) and why that matters specifically for O2C: never ask it to calculate days-past-due, compute that yourself and pass it in as a fact.
Read the full post, with all three live API tests and the code, on RevExOS →