How it uses Jev
At each generation step, Jev scores a question about the user's prompt and the reply emitted so far, choosing among an alphabet plus a stop option. A sampler draws the next symbol from the returned distribution and appends it.
What Jev decides
Illustrative Jev requestExample answers · not a recorded Jev response · Source ↗
Question 1 · next symbol
YOUR APP
INSTRUCTION
Append one option to answer_so_far. Choose the option that makes the result a valid prefix of a grammatical, factually correct English answer; choose STOP only when the question is complete.
STATE
Task/question, conversation if present and partial reply; configured lower26 alphabet.
JEV · CHOICE
- a
- b
- c
- d
- e
- f
- g
- h
- i
- j
- k
- l
- m
- n
- o
- p
- q
- r
- s
- t
- u
- v
- w
- x
- y
- z
- STOP
App workflow
- Build the next-symbol question
Combine the original prompt with the partial reply and present the configured symbol options and stop option.
- Sample a symbol
Use Jev's option probabilities with the selected sampling strategy to draw the next symbol.
- Continue or stop
Append the symbol and repeat until the stop option is selected.
Why it is interesting
It turns a classifier-style decision interface into a character or token sampler, with alternate alphabets and search strategies that make the tradeoffs inspectable.