jev-dsl

# jev-dsl

A Haskell DSL for agents that would rather have a question judged than
guessed. A packet of labelled questions is written once as an expression;
its type is inferred, it renders to the exact request JSON for
[TypeSafe's Jev](https://docs.typesafe.ai), and the answers come back under
the same labels as records read by field. A choice is consumed through
exhaustive labelled handlers or its carried payload, with a policy when
needed. The branch that runs is one the program wrote and carries the
payload it was offered with. No schema, no instance, no codec, no network.

**Early alpha.** The interface is still moving, and this repository is the
only place it has been used. See [Status](#status) for what is unsettled.

`Jev.Operators` is the whole surface, and the guide written for a model is
[docs/authoring.md](docs/authoring.md). `Jev.Core` is the same library
polymorphic over the JSON type, for an environment that cannot load aeson.
Every example below is compiled, by `test/Readme.hs` or as part of
`examples/Guard.hs`.

## What Jev is

[Jev](https://docs.typesafe.ai) is a hosted judgment model from TypeSafe.
It is not a chat model and there is no prompt. You send it a **state** —
whatever the situation is, as JSON — and a **map of labelled questions**
about that state, and it answers all of them in one call, as probabilities
rather than prose.

There are three kinds of question and nothing else:

| Kind | Asks | Comes back as |
|---|---|---|
| **Noul** | a proposition | one probability that it holds |
| **choice** | which one of these labelled alternatives applies | the winner, a mass for every alternative, and a confidence |
| **score** | where this falls on an ordered rubric of one to ten levels | a mass per level and an expectation |

That is the whole vocabu