diff --git a/Makefile b/Makefile
index 370aaa3..2453639 100644
--- a/Makefile
+++ b/Makefile
@@ -3,8 +3,8 @@
 # `make build`   — compile the Go binary into ./bin/mgpumanager.
 # `make test`    — go test ./...
 # `make run`     — run locally against ./config/consumers.yaml.
-# `make deploy`  — rsync binary + config + systemd unit to mRock,
-#                  reload systemd, restart the service.
+# `make deploy`  — rsync binary + config + user-unit to mRock and
+#                  (re)start it under `systemctl --user`.
 
 BIN := bin/mgpumanager
 PKG := ./cmd/mgpumanager
@@ -12,6 +12,7 @@ PKG := ./cmd/mgpumanager
 GO ?= go
 HOST ?= mrock
 REMOTE_DIR ?= /home/m/dev/mGPUmanager
+USER_UNIT_DIR ?= /home/m/.config/systemd/user
 
 .PHONY: build test run deploy clean
 
@@ -25,11 +26,13 @@ test:
 run: build
 	./$(BIN) --config config/consumers.yaml --log-level debug
 
+# Deploys to mRock as a user unit (systemd --user). User lingering must
+# be enabled on the target host: `sudo loginctl enable-linger m`.
 deploy: build
 	rsync -a --mkpath $(BIN) $(HOST):$(REMOTE_DIR)/$(BIN)
 	rsync -a --mkpath config/consumers.yaml $(HOST):$(REMOTE_DIR)/config/consumers.yaml
-	rsync -a --mkpath systemd/mgpumanager.service $(HOST):$(REMOTE_DIR)/systemd/mgpumanager.service
-	ssh $(HOST) "sudo cp $(REMOTE_DIR)/systemd/mgpumanager.service /etc/systemd/system/mgpumanager.service && sudo systemctl daemon-reload && sudo systemctl enable mgpumanager.service && sudo systemctl restart mgpumanager.service && sudo systemctl status mgpumanager.service --no-pager -l"
+	rsync -a --mkpath systemd/mgpumanager.service $(HOST):$(USER_UNIT_DIR)/mgpumanager.service
+	ssh $(HOST) "systemctl --user daemon-reload && systemctl --user enable mgpumanager.service && systemctl --user restart mgpumanager.service && systemctl --user status mgpumanager.service --no-pager -l"
 
 clean:
 	rm -rf bin
diff --git a/config/consumers.yaml b/config/consumers.yaml
index 4857c63..1792392 100644
--- a/config/consumers.yaml
+++ b/config/consumers.yaml
@@ -1,4 +1,4 @@
-listen: 127.0.0.1:8770
+listen: 0.0.0.0:8770
 
 gpu:
   total_mib: 16376        # RTX 4070 Ti SUPER
diff --git a/internal/scheduler/evicting.go b/internal/scheduler/evicting.go
index e0813df..3806709 100644
--- a/internal/scheduler/evicting.go
+++ b/internal/scheduler/evicting.go
@@ -70,13 +70,38 @@ func NewEvicting(cfg *config.Config, reg *registry.Registry, gpuPoller *gpu.Poll
 		lastUsed: make(map[string]time.Time, len(cfg.Consumers)),
 	}
 	for name, cons := range cfg.Consumers {
-		// Self-managed VRAM consumers (ollama) are always 'loaded' from
-		// the scheduler's perspective — we never evict them via HTTP.
-		e.loaded[name] = !cons.VRAMManaged || true
+		e.loaded[name] = initialLoaded(cons)
 	}
 	return e
 }
 
+// initialLoaded picks the believed-loaded state for a consumer at scheduler
+// startup. The rule:
+//
+//   - VRAM-managed (ollama): true — we never track or evict it.
+//   - Has a load route AND an unload route (mvoice): true — the consumer
+//     is set up to be controllable in both directions, and typically
+//     preloads on its own systemd-managed startup.
+//   - Has only an unload route, no load route (comfyui): false — lazy.
+//     FLUX isn't resident until the first /prompt; until that happens we
+//     don't account for its VRAM cost.
+//   - Has a systemd_unit but no HTTP routes (whisper-server): true — these
+//     are always-on services that load their model at process start.
+//   - Neither: true — fallback, assume it's there if the consumer is up.
+//
+// Getting this right matters for the eviction smoke test: if comfyui were
+// believed loaded at startup, ensureFits would short-circuit on the first
+// /v1/image request and never trigger eviction. (m/mGPUmanager#1 live deploy.)
+func initialLoaded(cons *config.Consumer) bool {
+	if cons.VRAMManaged {
+		return true
+	}
+	if cons.Load == nil && cons.Unload != nil {
+		return false
+	}
+	return true
+}
+
 // Run is the public Scheduler interface: ensure room + load + serialise.
 func (e *Evicting) Run(ctx context.Context, consumer string, fn Job) error {
 	if err := e.ensureFits(ctx, consumer); err != nil {
diff --git a/internal/scheduler/evicting_test.go b/internal/scheduler/evicting_test.go
index 153e01a..01dc839 100644
--- a/internal/scheduler/evicting_test.go
+++ b/internal/scheduler/evicting_test.go
@@ -118,6 +118,37 @@ func buildCfg(mvoiceURL, comfyURL string) *config.Config {
 	}
 }
 
+// TestInitialLoadedHeuristic pins the comfyui-isn't-preloaded rule down:
+// a consumer with Unload but no Load is lazy; everything else is assumed
+// resident at startup.
+func TestInitialLoadedHeuristic(t *testing.T) {
+	cases := []struct {
+		name string
+		cons *config.Consumer
+		want bool
+	}{
+		{"vram_managed (ollama)", &config.Consumer{VRAMManaged: true}, true},
+		{"load+unload (mvoice)", &config.Consumer{
+			Load:   &config.Route{Path: "/load"},
+			Unload: &config.Route{Path: "/unload"},
+		}, true},
+		{"unload only — lazy (comfyui)", &config.Consumer{
+			Unload: &config.Route{Path: "/api/free"},
+		}, false},
+		{"systemd unit only (whisper-server)", &config.Consumer{
+			SystemdUnit: "whisper-server.service",
+		}, true},
+		{"empty consumer", &config.Consumer{}, true},
+	}
+	for _, c := range cases {
+		t.Run(c.name, func(t *testing.T) {
+			if got := initialLoaded(c.cons); got != c.want {
+				t.Errorf("initialLoaded = %v, want %v", got, c.want)
+			}
+		})
+	}
+}
+
 // TestEvictingSkipsWhenAlreadyResident verifies the no-op fast path: a job
 // for an already-loaded consumer with plenty of free VRAM runs without any
 // unload call.
diff --git a/systemd/mgpumanager.service b/systemd/mgpumanager.service
index 2a07b53..d9b6c47 100644
--- a/systemd/mgpumanager.service
+++ b/systemd/mgpumanager.service
@@ -1,30 +1,15 @@
 [Unit]
-Description=mGPUmanager — GPU-Inference-Control-Plane for mRock
+Description=mGPUmanager — GPU-Inference-Control-Plane
 Documentation=https://mgit.msbls.de/m/mGPUmanager
-After=network-online.target
-Wants=network-online.target
+After=network.target
 
 [Service]
 Type=simple
-User=m
-Group=m
-WorkingDirectory=/home/m/dev/mGPUmanager
-ExecStart=/home/m/dev/mGPUmanager/bin/mgpumanager \
-    --config /home/m/dev/mGPUmanager/config/consumers.yaml \
-    --log-level info
+WorkingDirectory=%h/dev/mGPUmanager
+ExecStart=%h/dev/mGPUmanager/bin/mgpumanager --config %h/dev/mGPUmanager/config/consumers.yaml --log-level info
 Restart=on-failure
-RestartSec=3
+RestartSec=5
 TimeoutStopSec=10
 
-# Hardening — broker has no need for elevated capabilities.
-NoNewPrivileges=true
-PrivateTmp=true
-ProtectSystem=strict
-ProtectHome=read-only
-ReadWritePaths=/home/m/dev/mGPUmanager
-
-# The broker only proxies; nvidia-smi is the only GPU-touching call.
-PrivateDevices=false
-
 [Install]
-WantedBy=multi-user.target
+WantedBy=default.target