Jev playground

Drop-in OpenAI completions API

Point any OpenAI client at https://your-app.lovable.app/api/public/v1 with model jev-alphabet-001. Any API key works; there is nothing to protect here. /completions, /chat/completions and /models are implemented, with streaming and logprobs.

curl https://your-app.lovable.app/api/public/v1/completions \
  -H "Content-Type: application/json" \
  -d '{"model":"jev-alphabet-001","prompt":"The capital of France is","max_tokens":20,"logprobs":5}'

Every character is one Jev call: the options are the bare alphabet, with no dictionary behind them. On 60 held-out cuts of real prose it names the next character correctly 56.7% of the time, against 86.7% for a frontier model asked the same 60 questions. Everything highlighted is Jev’s own distribution; impossible options (a third repeat of a cycle, whitespace twice over, an early “I’m done”) are removed from the answer set rather than answered for it.