Live API calls · 15 October 2026

What we measured

Two travel chats, 20 labeled turns, using gpt-6-luna.

The snapshot total includes the model calls that create and update it. This is a small scripted test. See the full method, older-model tests, and other results.

Jev result: Ordinary snapshot updates missed preference changes. A change-only update with jev-latest matched 20/20 mock tool choices twice on new scripted chats, using 2.7% less model input than full-history Jev. On earlier longer chats it saved 17.5%. This prototype turns explicit changes into hard yes/no values. See the Jev comparison.

Use a score to make a decision

The model estimates a score. The rule is deterministic: the same score and threshold produce the same tool call.

when confidence(intent.book_flight) > 0.9:
  call search_flights()

Settle each field separately

Choose a confidence threshold for each decision field. When its current value is certain enough, save it and omit that field from the next update. Keep updating uncertain fields. Reopen a settled field when new input could change it.

intent = book_flight (0.97) → settled
prefers_direct = uncertain (0.58) → update next turn

In a separate gpt-6-luna test, settled fields matched 20/20 labeled turns twice and used about 10% fewer tokens: 5,828 vs 6,494, then 5,809 vs 6,458. At a 0.90 threshold, 54 of 60 numeric field updates were omitted. See the method and limits.

Try it locally

Build the project, then run an example:

npm install
npm run build
node dist/src/cli.js run examples/intent.bel --trace

With OPENAI_API_KEY in your environment or local .env file, try the chat modes:

npm run chatbot
CHAT_MODE=full-history npm run chatbot
CHAT_MODE=coded npm run chatbot

For benchmark commands and details, see the README.

Language basics

Define beliefs, then act when a score passes a threshold:

belief intent exclusive closed {
  book_flight: 0.82
  book_hotel: 0.12
  unknown: 0.06
}

when confidence(intent.book_flight) > 0.7:
  call search_flights()

See the full language reference for syntax and behavior.

Runtime API

Read state with getState() or confidence(path). Save and restore it with getBeliefSnapshot() and loadBeliefSnapshot(). Inspect decisions with getProvenance(), explainBelief(path), and --trace.

BeliefLang can infer with GPT or Jev. The integration and benchmark notes explain both adapters.