diff --git a/core/state/snapshot/difflayer.go b/core/state/snapshot/difflayer.go
index c7a65e2a4b3ef94154c084eb5b03806d31a2e00a..0f7a4223fb9af6fc4b5dabb206c70c395f40ff4e 100644
--- a/core/state/snapshot/difflayer.go
+++ b/core/state/snapshot/difflayer.go
@@ -22,8 +22,6 @@ import (
 	"sync"
 
 	"github.com/ethereum/go-ethereum/common"
-	"github.com/ethereum/go-ethereum/core/rawdb"
-	"github.com/ethereum/go-ethereum/ethdb"
 	"github.com/ethereum/go-ethereum/log"
 	"github.com/ethereum/go-ethereum/rlp"
 )
@@ -169,125 +167,6 @@ func (dl *diffLayer) Update(blockRoot common.Hash, accounts map[common.Hash][]by
 	return newDiffLayer(dl, dl.number+1, blockRoot, accounts, storage)
 }
 
-// Cap traverses downwards the diff tree until the number of allowed layers are
-// crossed. All diffs beyond the permitted number are flattened downwards. If
-// the layer limit is reached, memory cap is also enforced (but not before). The
-// block numbers for the disk layer and first diff layer are returned for GC.
-func (dl *diffLayer) Cap(layers int, memory uint64) (uint64, uint64) {
-	// Dive until we run out of layers or reach the persistent database
-	if layers > 2 {
-		// If we still have diff layers below, recurse
-		if parent, ok := dl.parent.(*diffLayer); ok {
-			return parent.Cap(layers-1, memory)
-		}
-		// Diff stack too shallow, return block numbers without modifications
-		return dl.parent.(*diskLayer).number, dl.number
-	}
-	// We're out of layers, flatten anything below, stopping if it's the disk or if
-	// the memory limit is not yet exceeded.
-	switch parent := dl.parent.(type) {
-	case *diskLayer:
-		return parent.number, dl.number
-	case *diffLayer:
-		// Flatten the parent into the grandparent. The flattening internally obtains a
-		// write lock on grandparent.
-		flattened := parent.flatten().(*diffLayer)
-
-		dl.lock.Lock()
-		defer dl.lock.Unlock()
-
-		dl.parent = flattened
-		if flattened.memory < memory {
-			diskNumber, _ := flattened.parent.Info()
-			return diskNumber, flattened.number
-		}
-	default:
-		panic(fmt.Sprintf("unknown data layer: %T", parent))
-	}
-	// If the bottommost layer is larger than our memory cap, persist to disk
-	var (
-		parent = dl.parent.(*diffLayer)
-		base   = parent.parent.(*diskLayer)
-		batch  = base.db.NewBatch()
-	)
-	parent.lock.RLock()
-	defer parent.lock.RUnlock()
-
-	// Start by temporarily deleting the current snapshot block marker. This
-	// ensures that in the case of a crash, the entire snapshot is invalidated.
-	rawdb.DeleteSnapshotBlock(batch)
-
-	// Mark the original base as stale as we're going to create a new wrapper
-	base.lock.Lock()
-	if base.stale {
-		panic("parent disk layer is stale") // we've committed into the same base from two children, boo
-	}
-	base.stale = true
-	base.lock.Unlock()
-
-	// Push all the accounts into the database
-	for hash, data := range parent.accountData {
-		if len(data) > 0 {
-			// Account was updated, push to disk
-			rawdb.WriteAccountSnapshot(batch, hash, data)
-			base.cache.Set(string(hash[:]), data)
-
-			if batch.ValueSize() > ethdb.IdealBatchSize {
-				if err := batch.Write(); err != nil {
-					log.Crit("Failed to write account snapshot", "err", err)
-				}
-				batch.Reset()
-			}
-		} else {
-			// Account was deleted, remove all storage slots too
-			rawdb.DeleteAccountSnapshot(batch, hash)
-			base.cache.Set(string(hash[:]), nil)
-
-			it := rawdb.IterateStorageSnapshots(base.db, hash)
-			for it.Next() {
-				if key := it.Key(); len(key) == 65 { // TODO(karalabe): Yuck, we should move this into the iterator
-					batch.Delete(key)
-					base.cache.Delete(string(key[1:]))
-				}
-			}
-			it.Release()
-		}
-	}
-	// Push all the storage slots into the database
-	for accountHash, storage := range parent.storageData {
-		for storageHash, data := range storage {
-			if len(data) > 0 {
-				rawdb.WriteStorageSnapshot(batch, accountHash, storageHash, data)
-				base.cache.Set(string(append(accountHash[:], storageHash[:]...)), data)
-			} else {
-				rawdb.DeleteStorageSnapshot(batch, accountHash, storageHash)
-				base.cache.Set(string(append(accountHash[:], storageHash[:]...)), nil)
-			}
-		}
-		if batch.ValueSize() > ethdb.IdealBatchSize {
-			if err := batch.Write(); err != nil {
-				log.Crit("Failed to write storage snapshot", "err", err)
-			}
-			batch.Reset()
-		}
-	}
-	// Update the snapshot block marker and write any remainder data
-	newBase := &diskLayer{
-		root:    parent.root,
-		number:  parent.number,
-		cache:   base.cache,
-		db:      base.db,
-		journal: base.journal,
-	}
-	rawdb.WriteSnapshotBlock(batch, newBase.number, newBase.root)
-	if err := batch.Write(); err != nil {
-		log.Crit("Failed to write leftover snapshot", "err", err)
-	}
-	dl.parent = newBase
-
-	return newBase.number, dl.number
-}
-
 // flatten pushes all data from this point downwards, flattening everything into
 // a single diff at the bottom. Since usually the lowermost diff is the largest,
 // the flattening bulds up from there in reverse.
diff --git a/core/state/snapshot/difflayer_test.go b/core/state/snapshot/difflayer_test.go
index 5a718c6171581bb3fb998fcefe45dcf27a91755e..5499f201618914fe89380e3af2e02f5e909bd2b1 100644
--- a/core/state/snapshot/difflayer_test.go
+++ b/core/state/snapshot/difflayer_test.go
@@ -18,15 +18,11 @@ package snapshot
 
 import (
 	"bytes"
-	"fmt"
 	"math/big"
 	"math/rand"
 	"testing"
-	"time"
 
-	"github.com/allegro/bigcache"
 	"github.com/ethereum/go-ethereum/common"
-	"github.com/ethereum/go-ethereum/core/rawdb"
 	"github.com/ethereum/go-ethereum/rlp"
 )
 
@@ -192,113 +188,12 @@ func TestInsertAndMerge(t *testing.T) {
 	}
 }
 
-// TestCapTree tests some functionality regarding capping/flattening
-func TestCapTree(t *testing.T) {
-
-	var (
-		storage = make(map[common.Hash]map[common.Hash][]byte)
-	)
-	setAccount := func(accKey string) map[common.Hash][]byte {
-		return map[common.Hash][]byte{
-			common.HexToHash(accKey): randomAccount(),
-		}
-	}
-	// the bottom-most layer, aside from the 'disk layer'
-	cache, _ := bigcache.NewBigCache(bigcache.Config{ // TODO(karalabe): dedup
-		Shards:             1,
-		LifeWindow:         time.Hour,
-		MaxEntriesInWindow: 1 * 1024,
-		MaxEntrySize:       1,
-		HardMaxCacheSize:   1,
-	})
-
-	base := &diskLayer{
-		journal: "",
-		db:      rawdb.NewMemoryDatabase(),
-		cache:   cache,
-		number:  0,
-		root:    common.HexToHash("0x01"),
-	}
-	// The lowest difflayer
-	a1 := base.Update(common.HexToHash("0xa1"), setAccount("0xa1"), storage)
-
-	a2 := a1.Update(common.HexToHash("0xa2"), setAccount("0xa2"), storage)
-	b2 := a1.Update(common.HexToHash("0xb2"), setAccount("0xb2"), storage)
-
-	a3 := a2.Update(common.HexToHash("0xa3"), setAccount("0xa3"), storage)
-	b3 := b2.Update(common.HexToHash("0xb3"), setAccount("0xb3"), storage)
-
-	checkExist := func(layer *diffLayer, key string) error {
-		accountKey := common.HexToHash(key)
-		data, _ := layer.Account(accountKey)
-		if data == nil {
-			return fmt.Errorf("expected %x to exist, got nil", accountKey)
-		}
-		return nil
-	}
-	shouldErr := func(layer *diffLayer, key string) error {
-		accountKey := common.HexToHash(key)
-		data, err := layer.Account(accountKey)
-		if err == nil {
-			return fmt.Errorf("expected error, got data %x", data)
-		}
-		return nil
-	}
-
-	// check basics
-	if err := checkExist(b3, "0xa1"); err != nil {
-		t.Error(err)
-	}
-	if err := checkExist(b3, "0xb2"); err != nil {
-		t.Error(err)
-	}
-	if err := checkExist(b3, "0xb3"); err != nil {
-		t.Error(err)
-	}
-	// Now, merge the a-chain
-	diskNum, diffNum := a3.Cap(0, 1024)
-	if diskNum != 0 {
-		t.Errorf("disk layer err, got %d exp %d", diskNum, 0)
-	}
-	if diffNum != 2 {
-		t.Errorf("diff layer err, got %d exp %d", diffNum, 2)
-	}
-	// At this point, a2 got merged into a1. Thus, a1 is now modified,
-	// and as a1 is the parent of b2, b2 should no longer be able to iterate into parent
-
-	// These should still be accessible
-	if err := checkExist(b3, "0xb2"); err != nil {
-		t.Error(err)
-	}
-	if err := checkExist(b3, "0xb3"); err != nil {
-		t.Error(err)
-	}
-	//b2ParentNum, _ := b2.parent.Info()
-	//if b2.parent.invalid == false
-	//	t.Errorf("err, exp parent to be invalid, got %v", b2.parent, b2ParentNum)
-	//}
-	// But these would need iteration into the modified parent:
-	if err := shouldErr(b3, "0xa1"); err != nil {
-		t.Error(err)
-	}
-	if err := shouldErr(b3, "0xa2"); err != nil {
-		t.Error(err)
-	}
-	if err := shouldErr(b3, "0xa3"); err != nil {
-		t.Error(err)
-	}
-}
-
 type emptyLayer struct{}
 
 func (emptyLayer) Update(blockRoot common.Hash, accounts map[common.Hash][]byte, storage map[common.Hash]map[common.Hash][]byte) *diffLayer {
 	panic("implement me")
 }
 
-func (emptyLayer) Cap(layers int, memory uint64) (uint64, uint64) {
-	panic("implement me")
-}
-
 func (emptyLayer) Journal() error {
 	panic("implement me")
 }
@@ -403,7 +298,6 @@ func BenchmarkSearchSlot(b *testing.B) {
 // Without sorting and tracking accountlist
 // BenchmarkFlatten-6   	     300	   5511511 ns/op
 func BenchmarkFlatten(b *testing.B) {
-
 	fill := func(parent snapshot, blocknum int) *diffLayer {
 		accounts := make(map[common.Hash][]byte)
 		storage := make(map[common.Hash]map[common.Hash][]byte)
diff --git a/core/state/snapshot/disklayer.go b/core/state/snapshot/disklayer.go
index a9839f01a8d822a9d021303b191ca2706e994aae..50321f1540e768dc02e217ac16229a5bfe803810 100644
--- a/core/state/snapshot/disklayer.go
+++ b/core/state/snapshot/disklayer.go
@@ -126,12 +126,6 @@ func (dl *diskLayer) Update(blockHash common.Hash, accounts map[common.Hash][]by
 	return newDiffLayer(dl, dl.number+1, blockHash, accounts, storage)
 }
 
-// Cap traverses downwards the diff tree until the number of allowed layers are
-// crossed. All diffs beyond the permitted number are flattened downwards.
-func (dl *diskLayer) Cap(layers int, memory uint64) (uint64, uint64) {
-	return dl.number, dl.number
-}
-
 // Journal commits an entire diff hierarchy to disk into a single journal file.
 func (dl *diskLayer) Journal() error {
 	// There's no journalling a disk layer
diff --git a/core/state/snapshot/snapshot.go b/core/state/snapshot/snapshot.go
index 6a21d57dcb36c725ff34f303b48a2ddcd57870b3..a181789779710b5b888aa25c81ab355bc00435da 100644
--- a/core/state/snapshot/snapshot.go
+++ b/core/state/snapshot/snapshot.go
@@ -73,11 +73,6 @@ type snapshot interface {
 	// copying everything.
 	Update(blockRoot common.Hash, accounts map[common.Hash][]byte, storage map[common.Hash]map[common.Hash][]byte) *diffLayer
 
-	// Cap traverses downwards the diff tree until the number of allowed layers are
-	// crossed. All diffs beyond the permitted number are flattened downwards. The
-	// block numbers for the disk layer and first diff layer are returned for GC.
-	Cap(layers int, memory uint64) (uint64, uint64)
-
 	// Journal commits an entire diff hierarchy to disk into a single journal file.
 	// This is meant to be used during shutdown to persist the snapshot without
 	// flattening everything down (bad for reorgs).
@@ -169,11 +164,56 @@ func (st *SnapshotTree) Cap(blockRoot common.Hash, layers int, memory uint64) er
 	if snap == nil {
 		return fmt.Errorf("snapshot [%#x] missing", blockRoot)
 	}
+	diff, ok := snap.(*diffLayer)
+	if !ok {
+		return fmt.Errorf("snapshot [%#x] is base layer", blockRoot)
+	}
 	// Run the internal capping and discard all stale layers
 	st.lock.Lock()
 	defer st.lock.Unlock()
 
-	diskNumber, diffNumber := snap.Cap(layers, memory)
+	var (
+		diskNumber uint64
+		diffNumber uint64
+	)
+	// Flattening the bottom-most diff layer requires special casing since there's
+	// no child to rewire to the grandparent. In that case we can fake a temporary
+	// child for the capping and then remove it.
+	switch layers {
+	case 0:
+		// If full commit was requested, flatten the diffs and merge onto disk
+		diff.lock.RLock()
+		base := diffToDisk(diff.flatten().(*diffLayer))
+		diff.lock.RUnlock()
+
+		st.layers[base.root] = base
+		diskNumber, diffNumber = base.number, base.number
+
+	case 1:
+		// If full flattening was requested, flatten the diffs but only merge if the
+		// memory limit was reached
+		var (
+			bottom *diffLayer
+			base   *diskLayer
+		)
+		diff.lock.RLock()
+		bottom = diff.flatten().(*diffLayer)
+		if bottom.memory >= memory {
+			base = diffToDisk(bottom)
+		}
+		diff.lock.RUnlock()
+
+		if base != nil {
+			st.layers[base.root] = base
+			diskNumber, diffNumber = base.number, base.number
+		} else {
+			st.layers[bottom.root] = bottom
+			diskNumber, diffNumber = bottom.parent.(*diskLayer).number, bottom.number
+		}
+
+	default:
+		diskNumber, diffNumber = st.cap(diff, layers, memory)
+	}
 	for root, snap := range st.layers {
 		if number, _ := snap.Info(); number != diskNumber && number < diffNumber {
 			delete(st.layers, root)
@@ -182,6 +222,135 @@ func (st *SnapshotTree) Cap(blockRoot common.Hash, layers int, memory uint64) er
 	return nil
 }
 
+// cap traverses downwards the diff tree until the number of allowed layers are
+// crossed. All diffs beyond the permitted number are flattened downwards. If
+// the layer limit is reached, memory cap is also enforced (but not before). The
+// block numbers for the disk layer and first diff layer are returned for GC.
+func (st *SnapshotTree) cap(diff *diffLayer, layers int, memory uint64) (uint64, uint64) {
+	// Dive until we run out of layers or reach the persistent database
+	if layers > 2 {
+		// If we still have diff layers below, recurse
+		if parent, ok := diff.parent.(*diffLayer); ok {
+			return st.cap(parent, layers-1, memory)
+		}
+		// Diff stack too shallow, return block numbers without modifications
+		return diff.parent.(*diskLayer).number, diff.number
+	}
+	// We're out of layers, flatten anything below, stopping if it's the disk or if
+	// the memory limit is not yet exceeded.
+	switch parent := diff.parent.(type) {
+	case *diskLayer:
+		return parent.number, diff.number
+
+	case *diffLayer:
+		// Flatten the parent into the grandparent. The flattening internally obtains a
+		// write lock on grandparent.
+		flattened := parent.flatten().(*diffLayer)
+		st.layers[flattened.root] = flattened
+
+		diff.lock.Lock()
+		defer diff.lock.Unlock()
+
+		diff.parent = flattened
+		if flattened.memory < memory {
+			diskNumber, _ := flattened.parent.Info()
+			return diskNumber, flattened.number
+		}
+	default:
+		panic(fmt.Sprintf("unknown data layer: %T", parent))
+	}
+	// If the bottom-most layer is larger than our memory cap, persist to disk
+	bottom := diff.parent.(*diffLayer)
+
+	bottom.lock.RLock()
+	base := diffToDisk(bottom)
+	bottom.lock.RUnlock()
+
+	st.layers[base.root] = base
+	diff.parent = base
+
+	return base.number, diff.number
+}
+
+// diffToDisk merges a bottom-most diff into the persistent disk layer underneath
+// it. The method will panic if called onto a non-bottom-most diff layer.
+func diffToDisk(bottom *diffLayer) *diskLayer {
+	var (
+		base  = bottom.parent.(*diskLayer)
+		batch = base.db.NewBatch()
+	)
+	// Start by temporarily deleting the current snapshot block marker. This
+	// ensures that in the case of a crash, the entire snapshot is invalidated.
+	rawdb.DeleteSnapshotBlock(batch)
+
+	// Mark the original base as stale as we're going to create a new wrapper
+	base.lock.Lock()
+	if base.stale {
+		panic("parent disk layer is stale") // we've committed into the same base from two children, boo
+	}
+	base.stale = true
+	base.lock.Unlock()
+
+	// Push all the accounts into the database
+	for hash, data := range bottom.accountData {
+		if len(data) > 0 {
+			// Account was updated, push to disk
+			rawdb.WriteAccountSnapshot(batch, hash, data)
+			base.cache.Set(string(hash[:]), data)
+
+			if batch.ValueSize() > ethdb.IdealBatchSize {
+				if err := batch.Write(); err != nil {
+					log.Crit("Failed to write account snapshot", "err", err)
+				}
+				batch.Reset()
+			}
+		} else {
+			// Account was deleted, remove all storage slots too
+			rawdb.DeleteAccountSnapshot(batch, hash)
+			base.cache.Set(string(hash[:]), nil)
+
+			it := rawdb.IterateStorageSnapshots(base.db, hash)
+			for it.Next() {
+				if key := it.Key(); len(key) == 65 { // TODO(karalabe): Yuck, we should move this into the iterator
+					batch.Delete(key)
+					base.cache.Delete(string(key[1:]))
+				}
+			}
+			it.Release()
+		}
+	}
+	// Push all the storage slots into the database
+	for accountHash, storage := range bottom.storageData {
+		for storageHash, data := range storage {
+			if len(data) > 0 {
+				rawdb.WriteStorageSnapshot(batch, accountHash, storageHash, data)
+				base.cache.Set(string(append(accountHash[:], storageHash[:]...)), data)
+			} else {
+				rawdb.DeleteStorageSnapshot(batch, accountHash, storageHash)
+				base.cache.Set(string(append(accountHash[:], storageHash[:]...)), nil)
+			}
+		}
+		if batch.ValueSize() > ethdb.IdealBatchSize {
+			if err := batch.Write(); err != nil {
+				log.Crit("Failed to write storage snapshot", "err", err)
+			}
+			batch.Reset()
+		}
+	}
+	// Update the snapshot block marker and write any remainder data
+	rawdb.WriteSnapshotBlock(batch, bottom.number, bottom.root)
+	if err := batch.Write(); err != nil {
+		log.Crit("Failed to write leftover snapshot", "err", err)
+	}
+	return &diskLayer{
+		root:    bottom.root,
+		number:  bottom.number,
+		cache:   base.cache,
+		db:      base.db,
+		journal: base.journal,
+	}
+}
+
 // Journal commits an entire diff hierarchy to disk into a single journal file.
 // This is meant to be used during shutdown to persist the snapshot without
 // flattening everything down (bad for reorgs).
diff --git a/core/state/snapshot/snapshot_test.go b/core/state/snapshot/snapshot_test.go
index 903bd4a6f630a6cf27d4754c3ff11acdd9c99244..ecd39bf3e81d741d27510a82d2de0a1ca8a4c798 100644
--- a/core/state/snapshot/snapshot_test.go
+++ b/core/state/snapshot/snapshot_test.go
@@ -15,3 +15,289 @@
 // along with the go-ethereum library. If not, see <http://www.gnu.org/licenses/>.
 
 package snapshot
+
+import (
+	"fmt"
+	"testing"
+	"time"
+
+	"github.com/allegro/bigcache"
+	"github.com/ethereum/go-ethereum/common"
+	"github.com/ethereum/go-ethereum/core/rawdb"
+)
+
+// Tests that if a disk layer becomes stale, no active external references will
+// be returned with junk data. This version of the test flattens every diff layer
+// to check internal corner case around the bottom-most memory accumulator.
+func TestDiskLayerExternalInvalidationFullFlatten(t *testing.T) {
+	// Create an empty base layer and a snapshot tree out of it
+	cache, _ := bigcache.NewBigCache(bigcache.DefaultConfig(time.Minute))
+	base := &diskLayer{
+		db:    rawdb.NewMemoryDatabase(),
+		root:  common.HexToHash("0x01"),
+		cache: cache,
+	}
+	snaps := &SnapshotTree{
+		layers: map[common.Hash]snapshot{
+			base.root: base,
+		},
+	}
+	// Retrieve a reference to the base and commit a diff on top
+	ref := snaps.Snapshot(base.root)
+
+	accounts := map[common.Hash][]byte{
+		common.HexToHash("0xa1"): randomAccount(),
+	}
+	storage := make(map[common.Hash]map[common.Hash][]byte)
+	if err := snaps.Update(common.HexToHash("0x02"), common.HexToHash("0x01"), accounts, storage); err != nil {
+		t.Fatalf("failed to create a diff layer: %v", err)
+	}
+	if n := len(snaps.layers); n != 2 {
+		t.Errorf("pre-cap layer count mismatch: have %d, want %d", n, 2)
+	}
+	// Commit the diff layer onto the disk and ensure it's persisted
+	if err := snaps.Cap(common.HexToHash("0x02"), 0, 0); err != nil {
+		t.Fatalf("failed to merge diff layer onto disk: %v", err)
+	}
+	// Since the base layer was modified, ensure that data retrievald on the external reference fail
+	if acc, err := ref.Account(common.HexToHash("0x01")); err != ErrSnapshotStale {
+		t.Errorf("stale reference returned account: %#x (err: %v)", acc, err)
+	}
+	if slot, err := ref.Storage(common.HexToHash("0xa1"), common.HexToHash("0xb1")); err != ErrSnapshotStale {
+		t.Errorf("stale reference returned storage slot: %#x (err: %v)", slot, err)
+	}
+	if n := len(snaps.layers); n != 1 {
+		t.Errorf("post-cap layer count mismatch: have %d, want %d", n, 1)
+		fmt.Println(snaps.layers)
+	}
+}
+
+// Tests that if a disk layer becomes stale, no active external references will
+// be returned with junk data. This version of the test retains the bottom diff
+// layer to check the usual mode of operation where the accumulator is retained.
+func TestDiskLayerExternalInvalidationPartialFlatten(t *testing.T) {
+	// Create an empty base layer and a snapshot tree out of it
+	cache, _ := bigcache.NewBigCache(bigcache.DefaultConfig(time.Minute))
+	base := &diskLayer{
+		db:    rawdb.NewMemoryDatabase(),
+		root:  common.HexToHash("0x01"),
+		cache: cache,
+	}
+	snaps := &SnapshotTree{
+		layers: map[common.Hash]snapshot{
+			base.root: base,
+		},
+	}
+	// Retrieve a reference to the base and commit two diffs on top
+	ref := snaps.Snapshot(base.root)
+
+	accounts := map[common.Hash][]byte{
+		common.HexToHash("0xa1"): randomAccount(),
+	}
+	storage := make(map[common.Hash]map[common.Hash][]byte)
+	if err := snaps.Update(common.HexToHash("0x02"), common.HexToHash("0x01"), accounts, storage); err != nil {
+		t.Fatalf("failed to create a diff layer: %v", err)
+	}
+	if err := snaps.Update(common.HexToHash("0x03"), common.HexToHash("0x02"), accounts, storage); err != nil {
+		t.Fatalf("failed to create a diff layer: %v", err)
+	}
+	if n := len(snaps.layers); n != 3 {
+		t.Errorf("pre-cap layer count mismatch: have %d, want %d", n, 3)
+	}
+	// Commit the diff layer onto the disk and ensure it's persisted
+	if err := snaps.Cap(common.HexToHash("0x03"), 2, 0); err != nil {
+		t.Fatalf("failed to merge diff layer onto disk: %v", err)
+	}
+	// Since the base layer was modified, ensure that data retrievald on the external reference fail
+	if acc, err := ref.Account(common.HexToHash("0x01")); err != ErrSnapshotStale {
+		t.Errorf("stale reference returned account: %#x (err: %v)", acc, err)
+	}
+	if slot, err := ref.Storage(common.HexToHash("0xa1"), common.HexToHash("0xb1")); err != ErrSnapshotStale {
+		t.Errorf("stale reference returned storage slot: %#x (err: %v)", slot, err)
+	}
+	if n := len(snaps.layers); n != 2 {
+		t.Errorf("post-cap layer count mismatch: have %d, want %d", n, 2)
+		fmt.Println(snaps.layers)
+	}
+}
+
+// Tests that if a diff layer becomes stale, no active external references will
+// be returned with junk data. This version of the test flattens every diff layer
+// to check internal corner case around the bottom-most memory accumulator.
+func TestDiffLayerExternalInvalidationFullFlatten(t *testing.T) {
+	// Create an empty base layer and a snapshot tree out of it
+	cache, _ := bigcache.NewBigCache(bigcache.DefaultConfig(time.Minute))
+	base := &diskLayer{
+		db:    rawdb.NewMemoryDatabase(),
+		root:  common.HexToHash("0x01"),
+		cache: cache,
+	}
+	snaps := &SnapshotTree{
+		layers: map[common.Hash]snapshot{
+			base.root: base,
+		},
+	}
+	// Commit two diffs on top and retrieve a reference to the bottommost
+	accounts := map[common.Hash][]byte{
+		common.HexToHash("0xa1"): randomAccount(),
+	}
+	storage := make(map[common.Hash]map[common.Hash][]byte)
+	if err := snaps.Update(common.HexToHash("0x02"), common.HexToHash("0x01"), accounts, storage); err != nil {
+		t.Fatalf("failed to create a diff layer: %v", err)
+	}
+	if err := snaps.Update(common.HexToHash("0x03"), common.HexToHash("0x02"), accounts, storage); err != nil {
+		t.Fatalf("failed to create a diff layer: %v", err)
+	}
+	if n := len(snaps.layers); n != 3 {
+		t.Errorf("pre-cap layer count mismatch: have %d, want %d", n, 3)
+	}
+	ref := snaps.Snapshot(common.HexToHash("0x02"))
+
+	// Flatten the diff layer into the bottom accumulator
+	if err := snaps.Cap(common.HexToHash("0x03"), 1, 1024*1024); err != nil {
+		t.Fatalf("failed to flatten diff layer into accumulator: %v", err)
+	}
+	// Since the accumulator diff layer was modified, ensure that data retrievald on the external reference fail
+	if acc, err := ref.Account(common.HexToHash("0x01")); err != ErrSnapshotStale {
+		t.Errorf("stale reference returned account: %#x (err: %v)", acc, err)
+	}
+	if slot, err := ref.Storage(common.HexToHash("0xa1"), common.HexToHash("0xb1")); err != ErrSnapshotStale {
+		t.Errorf("stale reference returned storage slot: %#x (err: %v)", slot, err)
+	}
+	if n := len(snaps.layers); n != 2 {
+		t.Errorf("post-cap layer count mismatch: have %d, want %d", n, 2)
+		fmt.Println(snaps.layers)
+	}
+}
+
+// Tests that if a diff layer becomes stale, no active external references will
+// be returned with junk data. This version of the test retains the bottom diff
+// layer to check the usual mode of operation where the accumulator is retained.
+func TestDiffLayerExternalInvalidationPartialFlatten(t *testing.T) {
+	// Create an empty base layer and a snapshot tree out of it
+	cache, _ := bigcache.NewBigCache(bigcache.DefaultConfig(time.Minute))
+	base := &diskLayer{
+		db:    rawdb.NewMemoryDatabase(),
+		root:  common.HexToHash("0x01"),
+		cache: cache,
+	}
+	snaps := &SnapshotTree{
+		layers: map[common.Hash]snapshot{
+			base.root: base,
+		},
+	}
+	// Commit three diffs on top and retrieve a reference to the bottommost
+	accounts := map[common.Hash][]byte{
+		common.HexToHash("0xa1"): randomAccount(),
+	}
+	storage := make(map[common.Hash]map[common.Hash][]byte)
+	if err := snaps.Update(common.HexToHash("0x02"), common.HexToHash("0x01"), accounts, storage); err != nil {
+		t.Fatalf("failed to create a diff layer: %v", err)
+	}
+	if err := snaps.Update(common.HexToHash("0x03"), common.HexToHash("0x02"), accounts, storage); err != nil {
+		t.Fatalf("failed to create a diff layer: %v", err)
+	}
+	if err := snaps.Update(common.HexToHash("0x04"), common.HexToHash("0x03"), accounts, storage); err != nil {
+		t.Fatalf("failed to create a diff layer: %v", err)
+	}
+	if n := len(snaps.layers); n != 4 {
+		t.Errorf("pre-cap layer count mismatch: have %d, want %d", n, 4)
+	}
+	ref := snaps.Snapshot(common.HexToHash("0x02"))
+
+	// Flatten the diff layer into the bottom accumulator
+	if err := snaps.Cap(common.HexToHash("0x04"), 2, 1024*1024); err != nil {
+		t.Fatalf("failed to flatten diff layer into accumulator: %v", err)
+	}
+	// Since the accumulator diff layer was modified, ensure that data retrievald on the external reference fail
+	if acc, err := ref.Account(common.HexToHash("0x01")); err != ErrSnapshotStale {
+		t.Errorf("stale reference returned account: %#x (err: %v)", acc, err)
+	}
+	if slot, err := ref.Storage(common.HexToHash("0xa1"), common.HexToHash("0xb1")); err != ErrSnapshotStale {
+		t.Errorf("stale reference returned storage slot: %#x (err: %v)", slot, err)
+	}
+	if n := len(snaps.layers); n != 3 {
+		t.Errorf("post-cap layer count mismatch: have %d, want %d", n, 3)
+		fmt.Println(snaps.layers)
+	}
+}
+
+// TestPostCapBasicDataAccess tests some functionality regarding capping/flattening.
+func TestPostCapBasicDataAccess(t *testing.T) {
+	// setAccount is a helper to construct a random account entry and assign it to
+	// an account slot in a snapshot
+	setAccount := func(accKey string) map[common.Hash][]byte {
+		return map[common.Hash][]byte{
+			common.HexToHash(accKey): randomAccount(),
+		}
+	}
+	// Create a starting base layer and a snapshot tree out of it
+	cache, _ := bigcache.NewBigCache(bigcache.DefaultConfig(time.Minute))
+	base := &diskLayer{
+		db:    rawdb.NewMemoryDatabase(),
+		root:  common.HexToHash("0x01"),
+		cache: cache,
+	}
+	snaps := &SnapshotTree{
+		layers: map[common.Hash]snapshot{
+			base.root: base,
+		},
+	}
+	// The lowest difflayer
+	snaps.Update(common.HexToHash("0xa1"), common.HexToHash("0x01"), setAccount("0xa1"), nil)
+	snaps.Update(common.HexToHash("0xa2"), common.HexToHash("0xa1"), setAccount("0xa2"), nil)
+	snaps.Update(common.HexToHash("0xb2"), common.HexToHash("0xa1"), setAccount("0xb2"), nil)
+
+	snaps.Update(common.HexToHash("0xa3"), common.HexToHash("0xa2"), setAccount("0xa3"), nil)
+	snaps.Update(common.HexToHash("0xb3"), common.HexToHash("0xb2"), setAccount("0xb3"), nil)
+
+	// checkExist verifies if an account exiss in a snapshot
+	checkExist := func(layer *diffLayer, key string) error {
+		if data, _ := layer.Account(common.HexToHash(key)); data == nil {
+			return fmt.Errorf("expected %x to exist, got nil", common.HexToHash(key))
+		}
+		return nil
+	}
+	// shouldErr checks that an account access errors as expected
+	shouldErr := func(layer *diffLayer, key string) error {
+		if data, err := layer.Account(common.HexToHash(key)); err == nil {
+			return fmt.Errorf("expected error, got data %x", data)
+		}
+		return nil
+	}
+	// check basics
+	snap := snaps.Snapshot(common.HexToHash("0xb3")).(*diffLayer)
+
+	if err := checkExist(snap, "0xa1"); err != nil {
+		t.Error(err)
+	}
+	if err := checkExist(snap, "0xb2"); err != nil {
+		t.Error(err)
+	}
+	if err := checkExist(snap, "0xb3"); err != nil {
+		t.Error(err)
+	}
+	// Now, merge the a-chain
+	snaps.Cap(common.HexToHash("0xa3"), 0, 1024)
+
+	// At this point, a2 got merged into a1. Thus, a1 is now modified, and as a1 is
+	// the parent of b2, b2 should no longer be able to iterate into parent.
+
+	// These should still be accessible
+	if err := checkExist(snap, "0xb2"); err != nil {
+		t.Error(err)
+	}
+	if err := checkExist(snap, "0xb3"); err != nil {
+		t.Error(err)
+	}
+	// But these would need iteration into the modified parent
+	if err := shouldErr(snap, "0xa1"); err != nil {
+		t.Error(err)
+	}
+	if err := shouldErr(snap, "0xa2"); err != nil {
+		t.Error(err)
+	}
+	if err := shouldErr(snap, "0xa3"); err != nil {
+		t.Error(err)
+	}
+}