A model that never writes a sentence
Decisions,
not text.
A lab for Jev, a model that answers typed questions instead of writing prose: send a state and a handful of questions, get a probability for every option and a confidence you can gate on. Six demos, each painting its answers the moment they arrive, and the written experiments behind them.
News with judgment
Real RSS headlines scored against a two-line profile of you. Cards climb the ranking as answers land: topic, clickbait, fact vs opinion.
Jarvis
One command in plain language, seventeen speculative questions at once. The code builds a plan and runs it on this Mac if confidence clears the bar.
App Store reviews
The hundred latest reviews of any app, sorted by severity as they're classified. Type, area, language, and whether the last update gets the blame.
Prompt pre-flight
Ten questions before an image provider sees the request: chat intent, policy category, block risk, evasion. Thresholds came out of 148 real prompts.
Line search, no embeddings
Every line of a pasted document becomes an option in one Choice. Ask in plain language and watch the heat map find the line, or admit the document doesn't say.
When it says 80%, is it right 80% of the time?
Thousands of human-labeled SMS, one yes/no question each, no training. The reliability diagram that tests the central claim: are the probabilities honest?
Experiments