Triage with Kai, reply with Zen

Route a support ticket with three typed decisions and no generated tokens, act only on the answers the model stands behind, then let Zen write the reply.

Three decisions, one generation. Which team, whether it is a bug and how urgent are bounded questions: Kai answers them as distributions. The reply is open text: Zen writes it.

Decide

cat > ticket.json <<'EOF'
{
  "model": "kai",
  "state": {"message": "I was charged twice for my March invoice and the second charge is still pending."},
  "questions": {
    "is_bug":  {"type": "noul", "instructions": "Is this a product defect?",
                "criteria": {"true": "a defect", "false": "working as intended"}},
    "team":    {"type": "choice", "instructions": "Which team should handle it?",
                "criteria": {"account": "logins and profiles",
                             "payments": "charges, invoices and refunds",
                             "product": "features and bugs"}},
    "urgency": {"type": "score", "instructions": "How urgent is it?",
                "criteria": ["later", "this week", "today", "now"]}
  }
}
EOF
d=$(curl -s "$DECISIONS/v1/decisions" -H "Content-Type: application/json" -d @ticket.json)

The baseline answered team: payments at 0.98, is_bug P(true) 0.31, and urgency with confidence 0.11 — four levels close to even. The full response is on the Kai page. usage.output_tokens is 0.

Act on what is settled

team=$(jq -r '.answers.team | if .confidence >= 0.5 then .choice else "review" end' <<<"$d")
urgency=$(jq -r '.answers.urgency
  | if .confidence >= 0.5 then .legend[(.probabilities | to_entries | max_by(.value).key)]
    else "review" end' <<<"$d")
echo "$team $urgency"    # payments review

confidence is 0 when the options are equally likely and 1 when one takes all the mass, so one threshold serves every question. Here team clears it and is routed; urgency does not, so it goes to a person instead of being guessed. Tune the threshold per question against your own outcomes, not by feel.

Reply

msg=$(jq -r .state.message ticket.json)
curl -s https://api.hanzo.ai/v1/chat/completions \
  -H "Authorization: Bearer $HANZO_API_KEY" \
  -H "Content-Type: application/json" \
  -d "$(jq -n --arg team "$team" --arg msg "$msg" '{
        model: "zen6",
        messages: [
          {role: "system", content: "You write for the \($team) team. Two sentences. Promise nothing you cannot see."},
          {role: "user", content: $msg}
        ]}')" \
  | jq -r '.choices[0].message.content'

What this buys. The routing questions cost one forward pass of a small encoder, return the same answer for the same ticket, and carry a hash of the weights and of the state they read. Generation is spent once, on the reply, where writing is the work.

Triage with Kai, reply with Zen

Decide

Act on what is settled

Reply

On this page