TypeSafe Jev for OTP.

Questions are data, the reply is a plain map, and Jev.Server lets a GenServer talk to Jev by replying:

def handle_call({:labels, issue}, from, s) do
  {:reply, {from, issue,
     kind: {"What kind of issue?", %{bug: nil, feature: nil, other: nil}},
     security: "Is this a vulnerability?"}, s}
end

def handle_answer(%{security: p}, from, s) when p > 0.5, do: done(from, [:security], s)
def handle_answer(%{kind: k, confidence: %{kind: c}}, from, s) when c > 0.85, do: done(from, [k], s)
def handle_answer(%{kind: k}, from, s), do: done(from, [k, :"needs-triage"], s)

This module holds the pure half: questions/1 normalizes shorthands into Jev.Noul, Jev.Choice, and Jev.Score structs, and reply/3 turns a decoded API response into the reply map. Jev.Wire describes the response body, and Jev.HTTP.post/3 is the transport.

The reply map

%{
  kind: :bug,            # Choice → label atom
  severity: 2.4,         # Score  → expected level, float
  security: 0.03,        # Noul   → probability of yes
  confidence: %{kind: 0.91, severity: 0.62},
  probabilities: %{kind: %{bug: 0.93, feature: 0.04, other: 0.03},
                   severity: %{0 => 0.1, 1 => 0.1, 2 => 0.2, 3 => 0.6}},
  usage: %{input_tokens: 812, output_tokens: 0, cost: 3.4e-5},
  model: "jev-1.13.0"    # the concrete model that answered, even when you asked for jev-latest
}

confidence, probabilities, usage, and model are reserved question names.

Other models

The wire format is served by open decision models as well as by TypeSafe, and reply/3 accepts what they send: usage may be missing, unknown fields are ignored, and when an answer carries probabilities but no confidence, confidence is computed the way TypeSafe defines it, as the top probability normalized over the number of options, (top - 1/k) / (1 - 1/k). A body that does not fit Jev.Wire at all is a JSONCodec.Error naming the field. Servers are calibrated differently, so a threshold tuned on one model is a starting point on another, not a guarantee.