Jev in practicewhat to ask, what to keep in code

Stop asking an LLM
for a yes/no.

Jev doesn’t write. You send it a state and typed questions; it returns a probability per question in ~300 ms. Your code decides.

Every number on this page is Jev’s own; the plan text is from a language model and labelled as such.

01The shape of one request

One request: a state and three typed questions in, three numbers out, and your code decides.

02One live question

Change the sentence. Watch the number.

▪ task_done · noul · live

One live Jev question. The task is fixed; edit the observation and Jev re-answers task_done — a noul, P(yes) — in ~300 ms.

taskBook a table for two at 7pm tonight.

observation

task_done—

the question we send

{
  "state": {
    "task": "Book a table for two at 7pm tonight.",
    "observation": "Reservation confirmed — 2 guests, 7:00 PM, Table 12."
  },
  "questions": {
    "task_done": {
      "type": "noul",
      "instructions": {
        "what": "Does `observation` confirm that the goal in `task` was achieved?",
        "not_for": "Partial progress, a different time or party size, or a promise to book later."
      },
      "criteria": {
        "true": "The observation confirms the task exactly as stated.",
        "false": "It does not confirm the task as stated."
      }
    }
  }
}

0 = no · 1 = yes · 0.5 is “don’t know”, not “medium”

03Three ways to ask

noul

Does observation confirm the goal in task?

“Reservation confirmed — 2 guests, 7:00 PM, Table 12.”

choice

“7:00 PM shows unavailable; 7:30 PM selectable”

confidence 0.84 · other: always leave an exit

score

0 nothing done
1 started
2 one subgoal confirmed
3 all but final
4 fully confirmed

levels are situations, not adjectives; the score is the probability-weighted mean

Stop asking an LLMfor a yes/no.

Change the sentence. Watch the number.

Stop asking an LLM
for a yes/no.