jevons-talking

Jev is a decision model: it can't write text, only pick from options. This gives it a voice anyway, borrowing the idea from the word-prediction keyboards used in assistive communication. Each turn, Jev sees a menu of predicted next words, letters to narrow the predictions, punctuation, backspace and SPEAK, and picks one. Code does everything else.

Q: What is the answer to life, the universe, and everything?

 1 word: the        conf=0.43  |
 2 backspace        conf=0.23  | the
 3 word: it         conf=0.18  |
 4 word: is         conf=0.45  | it
 5 letter: 4        conf=0.18  | it is
 6 letter: 2        conf=0.86  | it is 4…
 7 word: 42         conf=0.97  | it is 42…
 8 SPEAK            conf=0.80  | it is 42

A few more, unedited:

• Prediction is a deliberately dumb bigram model over WordNet's glosses and example sentences: no language model is doing the talking. Each menu offers 150 words — the 30 most common words, then the usual followers of the previous word. Typing a letter narrows the list to words with that prefix, which is how Jev reaches a word the predictor didn't think of.
• Memory. Every call is independent, so the state carries Jev's last 10 picks, and an option it backspaced over is dropped from that menu. Without this, Jev oscillates — adding a word, deleting it, adding it again — until it runs out of steps.
• Confidence. Each transcript in transcripts/ records every menu, pick, confidence and the full probability vector.

Needs Node 22+ and a TypeSafe API key in TYPESAFE_API_KEY (on macOS it's also read from the login keychain, service typesafe-jev).

npm install
npm run talk -- --q="Why is the sky blue?"

Each pick is one API call; a run stops at SPEAK or after --max-steps (default 100). --dry-run prints the first request without calling the API.