Interactive · uses a server-side key

Run a decision yourself

The four decision shapes from our benchmark, wired to the live model. Your browser holds no API key — it posts to this site, which forwards a fixed request shape and returns the answer with its token counts.

1 · Pick a decision shape

2 · Edit the state (the only thing you can change)

the question and the option list are fixed on the server79 / 1200

What you cannot change

The question and the options. They come from our server. If you could send an arbitrary prompt, this page would be a free gateway to a paid model on someone else's key.
The model. Pinned to jev-latest.

The key

Your browser never sees an API key. It posts to this site's own endpoint, which holds the key as a server secret and forwards a fixed request shape. The response you get back contains the answer, the token counts and the cost — nothing about the credential.

Limits

12 runs per visitor per 10 minutes
3,000 runs per day across everyone
input capped at 1,200 characters
no text from your input is stored

Limits exist because this runs on a real key with a real bill. When the daily budget is used up the page stops rather than spending more.

What gets logged

A hash of your IP address and a counter, so the limits can work. Not the input, not the answer, and not the raw address.

The exact request this page sends

One request, one decision

The payload is deliberately small. A state object holds whatever context the decision needs, and a questions object holds the question name plus the closed set of options. The answer comes back as a choice, a confidence, and a probability per option — no generated text anywhere in the path.

That is the whole interface. If you have used a chat completion to do this job, the difference worth noticing is that nothing here has to be parsed out of prose: the response is already the value your code branches on.

request

POST https://api.typesafe.ai/v1/systemone
{
  "model": "jev-latest",
  "state":     { "task": "<fixed by our server>", "input": "<what you typed>" },
  "questions": { "<name>": { "type": "choice", "criteria": { "<option>": "<description>", ... } } }
}

→ { "answers":  { "<name>": { "choice": "...", "confidence": 0.94, "probabilities": { ... } } },
    "usage":    { "input_tokens": 412, "output_tokens": 31 } }

Why the questions are fixed here

This page runs on our key, and the interesting risk is not the price of one call — it is becoming a free gateway to a paid model. So the server accepts a preset name and your state text, and nothing else. The question, the option list and the model are constants on our side.

The practical consequence is that you can explore how the model handles different inputs for a given decision shape, but you cannot repurpose the endpoint. That is a deliberate limit, and the same reasoning applies to any demo you build on a shared credential.