diff --git a/ansible/roles/honcho/README.md b/ansible/roles/honcho/README.md
index 6b9d567..b53bdaf 100644
--- a/ansible/roles/honcho/README.md
+++ b/ansible/roles/honcho/README.md
@@ -36,9 +36,42 @@ rootful Podman + Quadlet.
 4. Deploys a pgvector-enabled Postgres 16 container and waits for it to be
    ready.
 5. Deploys the Honcho API container (`honcho-api`) — FastAPI on :8000.
-6. Deploys the Honcho deriver worker container (`honcho-deriver`).
+6. Deploys the Honcho deriver worker container (`honcho-deriver`) — **stopped
+   and disabled by default** (see below).
 7. Verifies the API answers on `/docs`.
 
+## Deriver service (disabled by default)
+
+The `honcho-deriver` service is **created but not started** by default
+(`honcho_deriver_autostart: false`). This prevents autonomous token burn.
+
+**Why disabled:**  
+The deriver runs background reasoning loops that consume LLM tokens
+continuously, independent of the Hermes client-side `contextCadence` and
+`dialecticCadence` settings. Even with both cadences set to zero in
+`~/.hermes/honcho.json`, the deriver polls Postgres every second and fires
+Claude API calls autonomously.
+
+In production, this discovered behavior cost $10 USD overnight before the
+service was stopped.
+
+**Manual start (if needed for testing):**
+
+```bash
+# On lincoln.local.mk-labs.cloud as jarvis or with sudo:
+sudo systemctl start honcho-deriver.service
+
+# Check status:
+sudo systemctl status honcho-deriver.service
+
+# Stop again:
+sudo systemctl stop honcho-deriver.service
+```
+
+**To enable permanently:**  
+Override `honcho_deriver_autostart: true` in `group_vars` or inventory, then
+re-run the playbook.
+
 ## Required vault entries
 
 Add to `group_vars/all/vault` (ansible-vault encrypted):