diff --git a/spinifex/daemon/bedrock_deps.go b/spinifex/daemon/bedrock_deps.go
index 6199b9036..a2346acd4 100644
--- a/spinifex/daemon/bedrock_deps.go
+++ b/spinifex/daemon/bedrock_deps.go
@@ -70,7 +70,7 @@ func (d *Daemon) buildBedrockServiceDeps() handlers_bedrock.ServiceDeps {
 		Replicas: clusterSize,
 	}
 	if d.gpuManager != nil {
-		deps.GPU = &daemonGPUSnapshotter{mgr: d.gpuManager}
+		deps.GPU = &daemonGPUSnapshotter{d: d}
 	}
 	return deps
 }
@@ -78,10 +78,20 @@ func (d *Daemon) buildBedrockServiceDeps() handlers_bedrock.ServiceDeps {
 // daemonGPUSnapshotter adapts *gpu.Manager to handlers_bedrock's Snapshot-only
 // capacity-check surface, keeping the gpu package out of that handler package
 // the same way daemonGPUClaimer keeps it out of handlers_ec2_instance.
+//
+// It holds the daemon rather than the manager because applyGPUConfig swaps
+// d.gpuManager on a passthrough toggle. Capturing the pointer here would leave
+// this reading an orphaned pool that sees no claims and so admits everything.
 type daemonGPUSnapshotter struct {
-	mgr *gpu.Manager
+	d *Daemon
 }
 
 func (g *daemonGPUSnapshotter) Snapshot() []gpu.PoolEntry {
-	return g.mgr.Snapshot()
+	g.d.mu.Lock()
+	mgr := g.d.gpuManager
+	g.d.mu.Unlock()
+	if mgr == nil {
+		return nil
+	}
+	return mgr.Snapshot()
 }
diff --git a/spinifex/handlers/bedrock/capacity.go b/spinifex/handlers/bedrock/capacity.go
index 8a8c1e41d..c78f65ede 100644
--- a/spinifex/handlers/bedrock/capacity.go
+++ b/spinifex/handlers/bedrock/capacity.go
@@ -15,12 +15,16 @@ type gpuSnapshotter interface {
 	Snapshot() []gpu.PoolEntry
 }
 
-// checkCapacity reports whether the free pool (whole devices or MIG slices
-// with Available=true) has at least one entry with minVRAMMiB free. It is a
-// fast pre-admission check only: the deeper claim/allocate decision still
-// happens inside the shared RunInstances path when the VM actually launches,
-// so a device can still be lost to a racing launch between this check and
-// that claim — the guest fails to boot in that rare case, not silently OOMs.
+// checkCapacity reports whether the free pool (whole devices or MIG slices)
+// has at least one entry with minVRAMMiB free. It is a fast pre-admission
+// check only: the deeper claim/allocate decision still happens inside the
+// shared RunInstances path when the VM actually launches, so a device can
+// still be lost to a racing launch between this check and that claim — the
+// guest fails to boot in that rare case, not silently OOMs.
+//
+// Free means both axes the pool tracks: Available is device health, which
+// Release clears when a vfio rebind fails, while InstanceID is occupancy.
+// Claim tests both, and so must this — a claimed device keeps Available=true.
 func checkCapacity(snapshot []gpu.PoolEntry, minVRAMMiB int) error {
 	if minVRAMMiB <= 0 {
 		return fmt.Errorf("bedrock: invalid MinVRAMMiB %d", minVRAMMiB)
@@ -31,7 +35,7 @@ func checkCapacity(snapshot []gpu.PoolEntry, minVRAMMiB int) error {
 	var unknownVRAM *gpu.PoolEntry
 	for i := range snapshot {
 		entry := &snapshot[i]
-		if !entry.Available {
+		if !entry.Available || entry.InstanceID != "" {
 			continue
 		}
 		memMiB := entry.Device.MemoryMiB
diff --git a/spinifex/handlers/bedrock/capacity_test.go b/spinifex/handlers/bedrock/capacity_test.go
index 42e84d90f..b27c9c79e 100644
--- a/spinifex/handlers/bedrock/capacity_test.go
+++ b/spinifex/handlers/bedrock/capacity_test.go
@@ -31,13 +31,40 @@ func TestCheckCapacity(t *testing.T) {
 			minVRAM: 5120,
 		},
 		{
-			name: "device exists but not available (already claimed)",
+			name: "device unhealthy (rebind failed, Available cleared)",
 			snapshot: []gpu.PoolEntry{
 				{Device: gpu.GPUDevice{MemoryMiB: 8192}, Available: false},
 			},
 			minVRAM: 5120,
 			wantErr: true,
 		},
+		{
+			// The shape Claim actually leaves behind: it records InstanceID and
+			// leaves Available alone, so occupancy is invisible to a check that
+			// reads Available on its own.
+			name: "whole device claimed by a running instance",
+			snapshot: []gpu.PoolEntry{
+				{Device: gpu.GPUDevice{MemoryMiB: 8192}, Available: true, InstanceID: "i-abc"},
+			},
+			minVRAM: 5120,
+			wantErr: true,
+		},
+		{
+			name: "MIG slice claimed by a running instance",
+			snapshot: []gpu.PoolEntry{
+				{MIGInstance: &gpu.MIGInstance{Profile: gpu.MIGProfile{MemoryMiB: 10240}}, Available: true, InstanceID: "i-abc"},
+			},
+			minVRAM: 5120,
+			wantErr: true,
+		},
+		{
+			name: "a claimed device does not hide a free sibling",
+			snapshot: []gpu.PoolEntry{
+				{Device: gpu.GPUDevice{MemoryMiB: 16384}, Available: true, InstanceID: "i-abc"},
+				{Device: gpu.GPUDevice{MemoryMiB: 8192}, Available: true},
+			},
+			minVRAM: 5120,
+		},
 		{
 			name: "available device too small",
 			snapshot: []gpu.PoolEntry{
@@ -79,6 +106,34 @@ func TestCheckCapacity(t *testing.T) {
 	}
 }
 
+// TestCheckCapacity_ClaimedDeviceIsNotFree pins the defect that kept Ochre's
+// LRU eviction dormant: Claim records InstanceID and deliberately leaves
+// Available set, because Available tracks device health and Release clears it
+// only when a vfio rebind fails. Reading Available alone let a GPU held by a
+// running serving VM admit, so Ensure never took its capacity refusal and
+// evictForCapacity was never reached.
+func TestCheckCapacity_ClaimedDeviceIsNotFree(t *testing.T) {
+	pool := []gpu.PoolEntry{
+		{Device: gpu.GPUDevice{PCIAddress: "0000:5e:00.0", MemoryMiB: 8188}, Available: true, InstanceID: "i-abc"},
+	}
+	require.Error(t, checkCapacity(pool, 7168), "a GPU held by a running instance must not admit")
+
+	pool[0].InstanceID = ""
+	assert.NoError(t, checkCapacity(pool, 7168), "the same device must admit once released")
+}
+
+// TestCheckCapacity_UnknownVRAM_IgnoredWhenClaimed proves occupancy is settled
+// before the 0 MiB data-gap branch. A claimed device is not a candidate at all,
+// so an operator gets plain exhaustion rather than being sent to chase a
+// missing VRAM figure on a card that is merely busy.
+func TestCheckCapacity_UnknownVRAM_IgnoredWhenClaimed(t *testing.T) {
+	err := checkCapacity([]gpu.PoolEntry{
+		{Device: gpu.GPUDevice{PCIAddress: "0000:03:00.0", MemoryMiB: 0}, Available: true, InstanceID: "i-abc"},
+	}, 5120)
+	require.Error(t, err)
+	assert.NotContains(t, err.Error(), "undiscoverable")
+}
+
 func TestCheckCapacity_ErrorCode(t *testing.T) {
 	err := checkCapacity(nil, 5120)
 	assert.Equal(t, awserrors.ErrorModelNotReadyException, awserrors.ValidErrorCodeFromError(err))