diff --git a/src/psyc/cockpit/inference.py b/src/psyc/cockpit/inference.py
new file mode 100644
index 0000000..0e35eef
--- /dev/null
+++ b/src/psyc/cockpit/inference.py
@@ -0,0 +1,52 @@
+"""Inference client — calls the psyc model server if it is running.
+
+The cockpit venv has no torch; the fine-tuned model only runs inside the CUDA
+container behind serve_model.py. This client reaches it over HTTP and degrades
+gracefully — if the server is down, callers get None and fall back to rules.
+"""
+
+from __future__ import annotations
+
+import json
+from typing import Optional
+
+import httpx
+
+from psyc import log
+from psyc.lines.train import SEVERITY_INSTRUCTION, severity_features
+from psyc.models import Case
+
+
+INFERENCE_URL = "http://127.0.0.1:8771"
+
+_log = log.get(__name__)
+
+
+def server_adapter(timeout: float = 2.0) -> Optional[str]:
+    """Return the adapter the server is running, or None if it is unreachable."""
+    try:
+        with httpx.Client(timeout=timeout) as client:
+            resp = client.get(f"{INFERENCE_URL}/healthz")
+            resp.raise_for_status()
+            return resp.json().get("adapter")
+    except httpx.HTTPError:
+        return None
+
+
+def model_severity(case: Case, timeout: float = 15.0) -> Optional[str]:
+    """Ask the live model to classify case severity. None if the server is down."""
+    payload = {
+        "instruction": SEVERITY_INSTRUCTION,
+        "input": json.dumps(severity_features(case), ensure_ascii=False),
+        "max_new_tokens": 16,
+    }
+    try:
+        with httpx.Client(timeout=timeout) as client:
+            resp = client.post(f"{INFERENCE_URL}/infer", json=payload)
+            resp.raise_for_status()
+            output = str(resp.json().get("output", "")).strip().lower()
+    except httpx.HTTPError as exc:
+        _log.info("inference.unavailable", error=str(exc))
+        return None
+    _log.info("inference.severity", case_id=case.case_id, model_answer=output)
+    return output
diff --git a/src/psyc/cockpit/journey.py b/src/psyc/cockpit/journey.py
index 814ac25..e9274e2 100644
--- a/src/psyc/cockpit/journey.py
+++ b/src/psyc/cockpit/journey.py
@@ -7,10 +7,11 @@ a script. Beats that did not happen for a case are returned with occurred=False.
 
 from __future__ import annotations
 
-from typing import List
+from typing import List, Optional
 
 from pydantic import BaseModel, Field
 
+from psyc.cockpit import inference
 from psyc.lines import ledger as ledger_line
 from psyc.lines import route as route_line
 from psyc.lines import seal as seal_line
@@ -42,6 +43,7 @@ class Beat(BaseModel):
     occurred: bool
     caption: str
     facts: List[str] = Field(default_factory=list)
+    model_answer: Optional[str] = None  # live model prediction (Classifier bot only)
 
 
 def build_journey(case: Case) -> List[Beat]:
@@ -89,6 +91,7 @@ def _classified(case: Case) -> Beat:
         index=0, line="Classifyline", title="Classified", occurred=True,
         caption=f"Classifyline rated it {severity} severity, TLP:{c.tlp.value}, internal class {internal}.",
         facts=[f"incident type: {c.incident_type.value}"],
+        model_answer=inference.model_severity(case),
     )
 
 
diff --git a/src/psyc/cockpit/static/cockpit.css b/src/psyc/cockpit/static/cockpit.css
index 36fd00e..bf2f1fa 100644
--- a/src/psyc/cockpit/static/cockpit.css
+++ b/src/psyc/cockpit/static/cockpit.css
@@ -190,6 +190,23 @@ tr.sev-low .sev-badge { color: var(--muted); }
 .bot-facts { margin: 8px 0 0; padding-left: 16px; font-size: 12px; color: var(--muted); }
 .bot-facts li { margin: 2px 0; }
 
+/* live-model verdict on the Classifier bot */
+.bot-model {
+  margin: 10px 0 0; padding: 8px 10px; font-size: 13px;
+  border: 1px solid var(--border); border-radius: 6px; background: var(--navy-700);
+}
+.model-chip {
+  display: inline-block; font-size: 10px; font-weight: 700; letter-spacing: 0.6px;
+  text-transform: uppercase; padding: 2px 7px; border-radius: 4px; margin-right: 6px;
+  background: var(--cyan-700); color: var(--cyan-100); border: 1px solid var(--cyan-500);
+}
+.bot-model strong { color: var(--text); }
+.model-verdict { font-size: 12px; }
+.model-agree { border-color: color-mix(in oklab, var(--out-actioned) 45%, var(--border)); }
+.model-agree .model-verdict { color: var(--out-actioned); }
+.model-differ { border-color: color-mix(in oklab, var(--sev-medium) 45%, var(--border)); }
+.model-differ .model-verdict { color: var(--sev-medium); }
+
 .mesh.playing .bot-done .bot-working { animation: work-flash 1.5s ease both; animation-delay: calc(var(--i) * 0.95s); }
 @keyframes work-flash { 0%, 12% { opacity: 0; } 22%, 60% { opacity: 1; } 74%, 100% { opacity: 0; } }
 .mesh.playing .bot-done .bot-answer,
diff --git a/src/psyc/cockpit/templates/journey.html b/src/psyc/cockpit/templates/journey.html
index b163e55..7ccfb83 100644
--- a/src/psyc/cockpit/templates/journey.html
+++ b/src/psyc/cockpit/templates/journey.html
@@ -61,6 +61,16 @@
         {% if b.facts %}
         <ul class="bot-facts">{% for f in b.facts %}<li>{{ f }}</li>{% endfor %}</ul>
         {% endif %}
+        {% if b.model_answer %}
+        {% set rule_sev = case.classification.severity.value if case.classification.severity else '' %}
+        {% set agrees = b.model_answer == rule_sev %}
+        <p class="bot-model {{ 'model-agree' if agrees else 'model-differ' }}">
+          <span class="model-chip">⬡ psyc-v4 · live model</span>
+          severity: <strong>{{ b.model_answer | upper }}</strong>
+          {% if agrees %}<span class="model-verdict">✓ agrees with the rule</span>
+          {% else %}<span class="model-verdict">✗ differs — rule said {{ rule_sev | upper }}</span>{% endif %}
+        </p>
+        {% endif %}
       </div>
     </li>
     {% endfor %}
diff --git a/src/psyc/lines/train.py b/src/psyc/lines/train.py
index 3610c94..49c4d4c 100644
--- a/src/psyc/lines/train.py
+++ b/src/psyc/lines/train.py
@@ -98,10 +98,16 @@ def _ex_ioc_extraction(case: Case) -> Optional[Example]:
     )
 
 
-def _ex_severity_classification(case: Case) -> Optional[Example]:
-    if case.classification.severity is None:
-        return None
-    input_obj = {
+# Shared so the cockpit's live inference sends the exact prompt the model
+# trained on — keep _ex_severity_classification and the inference client in sync.
+SEVERITY_INSTRUCTION = (
+    "Classify the defensive severity of this case as one of: "
+    "low | medium | high | critical. Return only the label."
+)
+
+
+def severity_features(case: Case) -> Dict[str, object]:
+    return {
         "summary": case.summary,
         "source_type": case.source_type,
         "incident_type": case.classification.incident_type.value if case.classification.incident_type else None,
@@ -111,9 +117,14 @@ def _ex_severity_classification(case: Case) -> Optional[Example]:
         "url_count": len(case.observables.urls),
         "ip_count": len(case.observables.ips),
     }
+
+
+def _ex_severity_classification(case: Case) -> Optional[Example]:
+    if case.classification.severity is None:
+        return None
     return Example(
-        instruction="Classify the defensive severity of this case as one of: low | medium | high | critical. Return only the label.",
-        input=json.dumps(input_obj, ensure_ascii=False),
+        instruction=SEVERITY_INSTRUCTION,
+        input=json.dumps(severity_features(case), ensure_ascii=False),
         output=case.classification.severity.value,
         task="severity_classification",
         case_id=case.case_id,