Language model
Prompt plus the ticket
↓
Write tokens, one after another
↓
Parse the string into JSON
↓
{"refund": true} …or a paragraph
TypeSafe’s contrast. Output is a string. Code has to read it. The model can wander.
Trained classifier
Labeled tickets from last year
↓
Train once on a fixed label set
↓
New ticket in, one label out
↓
refund / not-refund
Usual other tool, not TypeSafe’s diagram. Change the question or the labels and you collect data and train again.
Jev
State: ticket, charges, policy
↓
Typed questions, written in language for this call
↓
All questions scored together
↓
refund 0.95 · billing · score 1.2
TypeSafe: possible answers defined in advance. No generated text. Code branches on the values.