diff --git a/PanTS-Demo/src/components/segmentation/RefineFlyout.tsx b/PanTS-Demo/src/components/segmentation/RefineFlyout.tsx
index 48944ea8..dfcbbe41 100644
--- a/PanTS-Demo/src/components/segmentation/RefineFlyout.tsx
+++ b/PanTS-Demo/src/components/segmentation/RefineFlyout.tsx
@@ -104,7 +104,7 @@ export default function RefineFlyout({ classLabel, refine, onApplied }: Props) {
 	return (
 		<div style={{ display: "flex", flexDirection: "column", gap: 8 }}>
 			<p className="atb-flyout-note">
-				The model redraws this class's outline from its current voxels, with no clicks. It only grows into unlabeled voxels, and Undo reverses it.
+				The model redraws this class's outline from its current voxels, with no clicks. It can add unlabeled voxels and trim this class's own, never touches another structure, and Undo reverses it.
 			</p>
 			{error && <p className="atb-flyout-note atb-flyout-note--error" role="alert">{error}</p>}
 			<MenuDivider />
diff --git a/PanTS-Demo/src/test/refineFlyout.test.tsx b/PanTS-Demo/src/test/refineFlyout.test.tsx
index f9b51771..6fff8555 100644
--- a/PanTS-Demo/src/test/refineFlyout.test.tsx
+++ b/PanTS-Demo/src/test/refineFlyout.test.tsx
@@ -1,6 +1,7 @@
 /**
- * The old standalone RefineFlyout no longer exists. Refinement now shares the
- * annotation ribbon's AI slot; this contract tests the current copy/state.
+ * Refinement shares the annotation ribbon's AI slot (the label and when it shows
+ * as "AI refine" are tested here); the flyout it opens is RefineFlyout.tsx, whose
+ * failure messages are tested in refineFlyoutErrors.test.tsx.
  */
 import { describe, expect, it } from "vitest";
 import { aiSlotCopy, deriveToolbarState, toolbarLayout } from "../components/viewer/annotationToolbarState";
diff --git a/docs/operations-reference.md b/docs/operations-reference.md
index d8064f64..fd00c48a 100644
--- a/docs/operations-reference.md
+++ b/docs/operations-reference.md
@@ -65,6 +65,60 @@ reachable server, a plain click or box falls back to the older region-grow tool,
 prompts only the model can answer (corrections, scribble, lasso) say the model
 is unavailable.
 
+#### Session slots
+
+The model server holds one "slot" (lease) per live prompt session and refuses a request
+when all slots are taken, which the viewer shows as "busy". `--max-sessions` sets the
+number of slots (default 3). One person can need two at once: the click, box, scribble
+and lasso tools keep one session per structure, and "Refine with model" opens a second,
+short-lived one. So:
+
+- Set `NNINTERACTIVE_MAX_SESSIONS` in `flask-server/.env` to the same number as the
+  server's `--max-sessions`, never higher. If the server has fewer, it refuses leases the
+  backend believes it has, and the backend logs a warning saying so.
+- Leave `NNINTERACTIVE_MAX_SESSIONS_PER_USER=2` (prompt session plus Refine).
+- With a single slot, Refine fails whenever the same person used a click tool in the last
+  ten minutes, and everyone else is refused while one person holds it.
+- A slot is freed when the viewer leaves the tool or the page closes, or after the idle
+  timeout (`NNINTERACTIVE_SESSION_TTL`, 600 s).
+
+GPU work is serialised, so with several people predicting at once the later ones wait.
+Measured on the lab host: a click or Refine on a real full-resolution scan answered in
+about 1 to 4 s alone, and three predictions started together on a synthetic volume
+finished about 4 s apart. Every slot also holds its CT in the model server's memory (a
+full-resolution scan is about 140 MB as float32), so check memory before raising the
+count a lot.
+
+#### Running the model server as a service
+
+On the lab host, run it as a systemd user service so it survives logouts and reboots
+(`loginctl enable-linger <user>` once). Save as `~/.config/systemd/user/nninteractive-bodymaps.service`:
+```
+[Unit]
+Description=nnInteractive model server for the BodyMaps website
+After=network.target
+
+[Service]
+ExecStart=/path/to/nninteractive-env/bin/nninteractive-server --model nnInteractive_v1.0 --device cuda --host 127.0.0.1 --port 1530 --no-torch-compile --max-sessions 3
+Restart=on-failure
+RestartSec=10
+StandardOutput=append:/path/to/nninteractive.log
+StandardError=append:/path/to/nninteractive.log
+
+[Install]
+WantedBy=default.target
+```
+Then `systemctl --user daemon-reload && systemctl --user enable --now nninteractive-bodymaps`,
+and point `NNINTERACTIVE_URL` at it. To change the slot count, edit `--max-sessions`,
+change `NNINTERACTIVE_MAX_SESSIONS` to match, `systemctl --user restart` the service, then
+restart the backend (`deploy/start_backend.sh`, at 0 running jobs). Restart when the log
+shows `0/N active`: anyone mid-edit loses their model session once, and the viewer
+restarts it from their current mask.
+
+To see what is happening: `GET /api/health` shows `interactive.leases` against
+`interactive.max` (this backend's own leases), and the model server's log prints
+`claimed session ... (k/N active)` and `released session ...` for every slot change.
+
 #### Create Conda Environment
 ```
 conda create -n PanTS_backend python=3.11
diff --git a/flask-server/.env.example b/flask-server/.env.example
index c1964276..34cfd680 100644
--- a/flask-server/.env.example
+++ b/flask-server/.env.example
@@ -51,8 +51,9 @@ PANTS_LABEL_SCHEMA=ipantsmini-v1
 # Preserve this fork's case/class/resolution sessions and baseline-injection
 # protocol with nninteractive-client 2.6.0. The model runs in a separate server.
 # NNINTERACTIVE_URL=http://127.0.0.1:1527
-# Maximum simultaneous leases held by this Flask process (default 2); the
-# model server permits 3 by default, leaving one for other processes.
+# Maximum simultaneous leases held by this Flask process (default 2). Keep it
+# no larger than the model server's --max-sessions (default 3), or the server
+# refuses leases this process thinks it can have and everyone sees "busy".
 # NNINTERACTIVE_MAX_SESSIONS=2
 # Maximum leases one signed-in reader may hold (default 2: prompt + Refine).
 # NNINTERACTIVE_MAX_SESSIONS_PER_USER=2
diff --git a/flask-server/api/api_blueprint.py b/flask-server/api/api_blueprint.py
index 7deee8d3..7f4a1c1c 100644
--- a/flask-server/api/api_blueprint.py
+++ b/flask-server/api/api_blueprint.py
@@ -3985,10 +3985,16 @@ def health():
     with _inference_jobs_lock:
         statuses = [job.get("status") for job in inference_jobs.values()]
     running, queued = statuses.count("running"), statuses.count("queued")
+    try:
+        from services import nninteractive_predictor as _nn
+        interactive = _nn.slot_stats()
+    except Exception:
+        interactive = None  # health must answer even if the model client is not installed
     resp = jsonify({
         "status": "ok" if mount["ok"] else "degraded",
         "mount": mount,
         "inference": {"running": running, "queued": queued, "active": running + queued},
+        "interactive": interactive,
         "threads": threading.active_count(),
     })
     resp.status_code = 200 if mount["ok"] else 503
diff --git a/flask-server/services/nninteractive_predictor.py b/flask-server/services/nninteractive_predictor.py
index 94e55c94..485ed1c8 100644
--- a/flask-server/services/nninteractive_predictor.py
+++ b/flask-server/services/nninteractive_predictor.py
@@ -415,6 +415,39 @@ def _evict_oldest_idle(predicate) -> bool:
     return False
 
 
+_LAST_LIMIT_WARNING = 0.0
+
+
+def _warn_if_model_server_refused_below_local_limit(exc: Exception) -> None:
+    """Say so when the model server is full but this backend thinks it has room.
+
+    This backend caps its own leases (NNINTERACTIVE_MAX_SESSIONS) and expects the
+    model server's --max-sessions to be at least as large. If the server refuses a
+    lease while this process holds fewer than its own limit, the two settings have
+    drifted apart (or another client is using the server), and every user sees
+    "busy". Logged at most once a minute. Caller holds _STORE_LOCK.
+    """
+    global _LAST_LIMIT_WARNING
+    if not _is_capacity_error(exc) or len(_SESSIONS) >= _MAX_SESSIONS:
+        return
+    now = time.time()
+    if now - _LAST_LIMIT_WARNING < 60:
+        return
+    _LAST_LIMIT_WARNING = now
+    print(
+        f"[nninteractive_predictor] WARNING: the model server refused a session while this backend holds "
+        f"only {len(_SESSIONS)} of its {_MAX_SESSIONS} (NNINTERACTIVE_MAX_SESSIONS). The model server's "
+        f"--max-sessions is lower than that, or another client is using it. Raise --max-sessions or lower "
+        f"NNINTERACTIVE_MAX_SESSIONS so they match."
+    )
+
+
+def slot_stats() -> dict:
+    """This backend's own model leases, for /api/health. Never raises or blocks on the model server."""
+    with _STORE_LOCK:
+        return {"leases": len(_SESSIONS), "max": _MAX_SESSIONS, "per_user": _MAX_SESSIONS_PER_USER}
+
+
 def session_is_active(
     case_id: str,
     segment_label: int | None,
@@ -470,7 +503,11 @@ def get_or_create_session(
                 raise PromptCapacityError("Interactive segmentation is at capacity; existing prompt sessions were preserved.")
 
         print(f"[nninteractive_predictor] creating session {key} (ct={ct.shape}, baseline={baseline_mask is not None})")
-        sess = _NNSession(ct, baseline_mask=baseline_mask)
+        try:
+            sess = _NNSession(ct, baseline_mask=baseline_mask)
+        except Exception as exc:
+            _warn_if_model_server_refused_below_local_limit(exc)
+            raise
         sess._bodymaps_owner = owner_key
         sess._bodymaps_token = session_token
         sess._bodymaps_case_id = case_id
diff --git a/flask-server/tests/unit/test_health_interactive.py b/flask-server/tests/unit/test_health_interactive.py
new file mode 100644
index 00000000..0484f53a
--- /dev/null
+++ b/flask-server/tests/unit/test_health_interactive.py
@@ -0,0 +1,29 @@
+"""/api/health reports this backend's model leases so "busy" can be told apart from "broken"."""
+from flask import Flask
+
+from services import nninteractive_predictor as predictor
+
+
+def _health_json():
+    from api.api_blueprint import health
+
+    app = Flask(__name__)
+    with app.test_request_context("/api/health"):
+        response = health()
+    return response.get_json()
+
+
+def test_health_includes_the_model_lease_counts():
+    body = _health_json()
+    assert body["interactive"]["max"] == predictor._MAX_SESSIONS
+    assert body["interactive"]["leases"] == len(predictor._SESSIONS)
+
+
+def test_health_still_answers_when_the_lease_stats_fail(monkeypatch):
+    def boom():
+        raise RuntimeError("model client missing")
+
+    monkeypatch.setattr(predictor, "slot_stats", boom)
+    body = _health_json()
+    assert body["interactive"] is None
+    assert body["status"] in {"ok", "degraded"}
diff --git a/flask-server/tests/unit/test_interactive_session_capacity.py b/flask-server/tests/unit/test_interactive_session_capacity.py
index 77176028..56d11d36 100644
--- a/flask-server/tests/unit/test_interactive_session_capacity.py
+++ b/flask-server/tests/unit/test_interactive_session_capacity.py
@@ -211,3 +211,55 @@ def close(self): self.closed = True
     assert FakeRemote.instances[1].calls[0] == ("seed", 1)
     assert result[0, 1, 1] == 1
     assert result[1, 1, 1] == 1
+
+
+def test_slot_stats_reports_this_backends_leases_and_limits(monkeypatch):
+    monkeypatch.setattr(predictor, "_NNSession", FakeSession)
+    ct = np.zeros((2, 2, 2), dtype=np.float32)
+    assert predictor.slot_stats() == {"leases": 0, "max": 2, "per_user": 2}
+    predictor.get_or_create_session("case-a", 1, "low", ct, session_token="a", owner="alice")
+    assert predictor.slot_stats()["leases"] == 1
+
+
+class RefusingSession:
+    def __init__(self, *_args, **_kwargs):
+        raise RuntimeError("server is at capacity")
+
+
+def test_warns_when_the_model_server_refuses_below_this_backends_own_limit(monkeypatch, capsys):
+    monkeypatch.setattr(predictor, "_NNSession", RefusingSession)
+    monkeypatch.setattr(predictor, "_LAST_LIMIT_WARNING", 0.0)
+    ct = np.zeros((2, 2, 2), dtype=np.float32)
+
+    with pytest.raises(RuntimeError, match="at capacity"):
+        predictor.get_or_create_session("case-a", 1, "low", ct, session_token="a", owner="alice")
+
+    out = capsys.readouterr().out
+    assert "--max-sessions" in out and "NNINTERACTIVE_MAX_SESSIONS" in out
+    assert "holds only 0 of its 2" in out
+
+
+def test_the_limit_warning_is_rate_limited(monkeypatch, capsys):
+    monkeypatch.setattr(predictor, "_NNSession", RefusingSession)
+    monkeypatch.setattr(predictor, "_LAST_LIMIT_WARNING", 0.0)
+    ct = np.zeros((2, 2, 2), dtype=np.float32)
+
+    for token in ("a", "b", "c"):
+        with pytest.raises(RuntimeError):
+            predictor.get_or_create_session("case-a", 1, "low", ct, session_token=token, owner="alice")
+
+    assert capsys.readouterr().out.count("WARNING: the model server refused") == 1
+
+
+def test_no_limit_warning_for_other_failures_or_when_the_local_pool_is_full(monkeypatch, capsys):
+    monkeypatch.setattr(predictor, "_LAST_LIMIT_WARNING", 0.0)
+    predictor._warn_if_model_server_refused_below_local_limit(RuntimeError("connection refused"))
+    assert "WARNING" not in capsys.readouterr().out
+
+    monkeypatch.setattr(predictor, "_NNSession", FakeSession)
+    ct = np.zeros((2, 2, 2), dtype=np.float32)
+    predictor.get_or_create_session("case-a", 1, "low", ct, session_token="a", owner="alice")
+    predictor.get_or_create_session("case-b", 1, "low", ct, session_token="b", owner="bob")
+    predictor._warn_if_model_server_refused_below_local_limit(RuntimeError("server is at capacity"))
+    assert "WARNING" not in capsys.readouterr().out
+