Small experiments.
Visible decisions.
Change the state. Inspect the distribution. Decide what your code should trust.
Real models only: Run decisions directly in your browser with Laya (LiteRT.js, mmBERT-base) or Kev (ONNX Runtime Web, 0.6B) with nothing sent anywhere, or hosted Jev with your key. No illustrative or synthetic mode. Try the full Playground.
Choose where decisions run
Local relay keeps keys in process memory only and forwards solely to TypeSafe, OpenAI or Anthropic. It is not offered on Pages. Kev still uses its loopback endpoint directly.
Laya runs on-device in this tab via LiteRT.js (mmBERT-base). Weights download once (~680 MB total); zero data leaves your browser.
Choice / Score: gate on returned confidence. Noul: positive if p ≥ threshold, negative if p ≤ 1 − threshold, otherwise review. This is a teaching policy, not a calibrated production threshold.
Decision & distribution
Ready. Select an experiment and run.
No action applied.
Request, response and provenance
Run an experiment to inspect the exact request and response. Keys are never included.
Make another experiment
A normal LLM writes a validated JSON specification; Jev or Kev makes its decisions. The fixed renderer supplies the controls and charts. This generator does not create arbitrary HTML, JavaScript, bespoke games or executable tools. Review its output before loading.
Generate sends this idea and the system prompt to the selected provider. It does not include your current experiment, decision key or prior results. Account access, billing and browser CORS must permit direct calls.
No generation yet. You can edit the checked example without a key.
Unsupported fields, code and invalid question schemas are refused. Strings resembling HTML remain plain text. Loading never calls an API.
Next experiment: unfamiliar labels, with and without examples · Why can a decision model be fast?