From first call to a safe training plan.

The mental model

Distill is a loop, not a one-click fine-tune. Evidence moves through six named stages, and each durable object gives both you and the copilot a reliable place to resume.

1. 1
   Capture

   Collect representative teacher calls and feedback.

2. 2
   Data

   Filter or import rows, then freeze a version.

3. 3
   Rewards

   Choose and test the signals training should optimize.

4. 4
   Plan & train

   Dry-run the exact recipe, approve it, then start live work.

5. 5
   Prove

   Read holdout evidence and inspect regressions.

6. 6
   Serve

   Promote from shadow to percentage to full traffic.

Create your first loop

Start in test mode with a descriptive name. Keep the returned loop_id; later datasets and sources can be scoped to it.

curl -X POST "https://<your-lizzy-host>/v1/distill/loops" \
+  -H "Authorization: Bearer $LIZZY_API_TOKEN" \
+  -H "Content-Type: application/json" \
+  -H "Idempotency-Key: quickstart-loop-001" \
+  -d '{"name":"Support quality","description":"Reduce cost while preserving resolution quality"}'

import os, requests

response = requests.post(
    "https://<your-lizzy-host>/v1/distill/loops",
    headers={
        "Authorization": f"Bearer {os.environ['LIZZY_API_TOKEN']}",
        "Idempotency-Key": "quickstart-loop-001",
    },
    json={"name": "Support quality", "description": "Preserve resolution quality"},
)
response.raise_for_status()
loop = response.json()

const response = await fetch("https://<your-lizzy-host>/v1/distill/loops", {
  method: "POST",
  headers: {
    Authorization: `Bearer ${process.env.LIZZY_API_TOKEN}`,
    "Content-Type": "application/json",
    "Idempotency-Key": "quickstart-loop-001",
  },
  body: JSON.stringify({ name: "Support quality", description: "Preserve resolution quality" }),
});
if (!response.ok) throw await response.json();
const loop = await response.json();

Work with the copilot

Open the copilot from any Distill page. Page context is attached to the conversation, so you can ask “what is blocking this run?” without pasting IDs. The same workflow appears in the full copilot page and the dock.

The copilot reports a phase, resource references, blockers, and a closed set of next actions. If a workflow stalls, inspect or resume that same workflow instead of opening a duplicate.

What happens next