diff --git a/crates/pdfs-fuse/src/filesystem.rs b/crates/pdfs-fuse/src/filesystem.rs
index fa22e7d..07013d7 100644
--- a/crates/pdfs-fuse/src/filesystem.rs
+++ b/crates/pdfs-fuse/src/filesystem.rs
@@ -1464,6 +1464,12 @@ impl ProtonFs {
             return;
         }
         if let Err(e) = self.core.ensure_children(parent) {
+            // The parent was looked up a moment ago, so ENOENT here means it
+            // left the tree in between. Log it: the caller only sees "no such
+            // file" on a create.
+            if e == Errno::ENOENT {
+                warn!(parent, name, "create: the parent folder is no longer known");
+            }
             reply.error(e);
             return;
         }
@@ -1472,6 +1478,10 @@ impl ProtonFs {
             match st.entries.get(&parent) {
                 Some(e) => e.uid.clone(),
                 None => {
+                    warn!(
+                        parent,
+                        name, "create: the parent folder left the tree during its listing"
+                    );
                     reply.error(Errno::ENOENT);
                     return;
                 }
diff --git a/docs/TESTING.md b/docs/TESTING.md
index 05db5ae..1233041 100644
--- a/docs/TESTING.md
+++ b/docs/TESTING.md
@@ -97,7 +97,8 @@ Besides the syscall contract, the suite covers:
 - **Throughput.** 64 MiB written and read sequentially, 1 MiB in 4 KiB writes,
   and metadata operations on 100 files. The rates are printed. The case fails
   below `PDFS_ACCEPTANCE_MIN_MIBPS` (default 10) or `PDFS_ACCEPTANCE_MIN_OPS`
-  (default 20 operations per second).
+  (default 20 operations per second, or 3 on a FUSE mount, where each create
+  and unlink waits for Drive before it returns).
 
 ### The mount is diffed against an ordinary filesystem
 
diff --git a/scripts/fuse-acceptance.py b/scripts/fuse-acceptance.py
index 6910dc9..4e76b51 100755
--- a/scripts/fuse-acceptance.py
+++ b/scripts/fuse-acceptance.py
@@ -1357,9 +1357,15 @@ def test_throughput(ctx: Context) -> None:
     reads of just-written data come from the cache, so neither should depend
     on the network. PDFS_ACCEPTANCE_MIN_MIBPS and PDFS_ACCEPTANCE_MIN_OPS move
     the floors; 0 turns one off.
+
+    Metadata is the exception. On a Drive mount each create and each unlink
+    waits for Drive before it returns (about 350 ms for a create), so 20
+    operations a second is out of reach there by design. The default floor for
+    a FUSE target is 3, which still catches a second round trip per operation.
     """
     min_rate = float(os.environ.get("PDFS_ACCEPTANCE_MIN_MIBPS", "10"))
-    min_ops = float(os.environ.get("PDFS_ACCEPTANCE_MIN_OPS", "20"))
+    default_ops = "3" if is_fuse(ctx.root) else "20"
+    min_ops = float(os.environ.get("PDFS_ACCEPTANCE_MIN_OPS", default_ops))
     root = ctx.root / "throughput"
     root.mkdir()
     chunk = pattern(MIB, "throughput")
@@ -3308,6 +3314,20 @@ def is_mountpoint(path: Path) -> bool:
     return os.path.ismount(path)
 
 
+def is_fuse(path: Path) -> bool:
+    """True when `path` lives on a FUSE mount, read from /proc/self/mountinfo."""
+    path = path.resolve()
+    best, fstype = -1, ""
+    with open("/proc/self/mountinfo", encoding="utf-8") as mounts:
+        for line in mounts:
+            fields = line.split()
+            # Octal escapes (\040 for a space) are how mountinfo writes odd names.
+            point = Path(fields[4].encode().decode("unicode_escape"))
+            if (path == point or point in path.parents) and len(point.parts) > best:
+                best, fstype = len(point.parts), fields[fields.index("-") + 1]
+    return fstype.startswith("fuse")
+
+
 def run_managed_matrix(
     paths: list[Path], reference: Run, args, janitor: Janitor | None = None
 ) -> None: