diff --git a/crates/cfb/README.md b/crates/cfb/README.md
index b0f0d8741..6b234b3ed 100644
--- a/crates/cfb/README.md
+++ b/crates/cfb/README.md
@@ -50,3 +50,6 @@ chain is contiguous. Sibling trees are balanced binary search trees in [MS-CFB]
 bytes) and multi-sector streams and up to 40 siblings; red-black invariants of every sibling tree;
 a 7.5 MB stream that needs DIFAT sectors; Unicode and case-insensitive names; truncation at every
 29 bytes and 1500 random mutations never panic and only return complete streams.
+The writer rejects a duplicate sibling name with an O(1) hash lookup on the upper-cased name (a scan over the
+siblings made n children cost O(n^2): 20 000 AAF markers took a minute); `adding_many_children_is_not_quadratic`
+and `duplicate_names_are_still_rejected` (`tests/writer_scale.rs`) cover it.
diff --git a/crates/cfb/src/write.rs b/crates/cfb/src/write.rs
index 923e1aa38..8a4ad4fbe 100644
--- a/crates/cfb/src/write.rs
+++ b/crates/cfb/src/write.rs
@@ -5,6 +5,8 @@
 //! when the FAT needs more than the 109 header slots. Every chain is contiguous. Sibling trees are
 //! balanced binary search trees coloured as valid red-black trees (§2.6.4).
 
+use std::collections::HashSet;
+
 use crate::{DIFSECT, ENDOFCHAIN, FATSECT, FREESECT, MINI_SECTOR_SIZE, MINI_STREAM_CUTOFF, NOSTREAM, Result, SIGNATURE, compare_names, validate_name};
 
 /// File format version.
@@ -44,6 +46,15 @@ struct Node {
     clsid: [u8; 16],
     data: Vec<u8>,
     children: Vec<usize>,
+    /// The children's names, upper-cased as sibling order compares them: an O(1) duplicate check.
+    /// (Scanning every sibling with `compare_names`, which allocates, made adding n children O(n^2):
+    /// 20 000 AAF markers took a minute.)
+    child_keys: HashSet<Vec<u16>>,
+}
+
+/// A name as sibling comparison sees it: its UTF-16 code units, upper-cased.
+fn name_key(name: &str) -> Vec<u16> {
+    name.encode_utf16().map(crate::upper).collect()
 }
 
 /// Builds a compound file in memory.
@@ -64,7 +75,17 @@ impl Writer {
     pub const ROOT: NodeId = NodeId(0);
 
     pub fn new(version: Version) -> Self {
-        Writer { version, nodes: vec![Node { name: "Root Entry".into(), kind: Kind::Root, clsid: [0; 16], data: Vec::new(), children: Vec::new() }] }
+        Writer {
+            version,
+            nodes: vec![Node {
+                name: "Root Entry".into(),
+                kind: Kind::Root,
+                clsid: [0; 16],
+                data: Vec::new(),
+                children: Vec::new(),
+                child_keys: HashSet::new(),
+            }],
+        }
     }
 
     fn add(&mut self, parent: NodeId, name: &str, kind: Kind, clsid: [u8; 16], data: Vec<u8>) -> Result<NodeId> {
@@ -73,12 +94,15 @@ impl Writer {
         if p.kind == Kind::Stream {
             return Err(crate::Error::Invalid(format!("{:?} is a stream, not a storage", p.name)));
         }
-        if p.children.iter().any(|&c| compare_names(&self.nodes[c].name, name).is_eq()) {
+        let key = name_key(name);
+        if p.child_keys.contains(&key) {
             return Err(crate::Error::BadName(format!("duplicate name {name}")));
         }
         let id = self.nodes.len();
-        self.nodes.push(Node { name: name.to_string(), kind, clsid, data, children: Vec::new() });
-        self.nodes[parent.0].children.push(id);
+        self.nodes.push(Node { name: name.to_string(), kind, clsid, data, children: Vec::new(), child_keys: HashSet::new() });
+        let p = &mut self.nodes[parent.0];
+        p.children.push(id);
+        p.child_keys.insert(key);
         Ok(NodeId(id))
     }
 
diff --git a/crates/cfb/tests/writer_scale.rs b/crates/cfb/tests/writer_scale.rs
new file mode 100644
index 000000000..47c40f9a2
--- /dev/null
+++ b/crates/cfb/tests/writer_scale.rs
@@ -0,0 +1,33 @@
+//! The writer scales: adding many children to one storage, and the duplicate-name rules that check keeps.
+
+use filmcraft_cfb::{Version, Writer};
+
+/// Adding a child compared its name with every sibling (allocating twice per comparison), so n children
+/// cost O(n^2): 20 000 AAF markers in one storage took about a minute. The duplicate check is a hash
+/// lookup now. 40 000 plain names take milliseconds now and about 18 s before (a minute for real marker
+/// objects), so the limit separates them by orders of magnitude on any machine.
+#[test]
+fn adding_many_children_is_not_quadratic() {
+    let mut w = Writer::new(Version::V3);
+    let start = std::time::Instant::now();
+    for i in 0..40_000 {
+        w.stream(Writer::ROOT, &format!("{i}"), vec![1]).expect("unique name");
+    }
+    assert!(start.elapsed() < std::time::Duration::from_secs(3), "{:?}", start.elapsed());
+}
+
+/// The faster duplicate check still means what sibling comparison means: case-insensitive, also beyond ASCII,
+/// per storage (the same name in another storage is fine), and a name that differs only in length is distinct.
+#[test]
+fn duplicate_names_are_still_rejected() {
+    let mut w = Writer::new(Version::V3);
+    let a = w.storage(Writer::ROOT, "Alpha", [0; 16]).unwrap();
+    assert!(w.stream(Writer::ROOT, "ALPHA", vec![]).is_err(), "case-insensitive");
+    assert!(w.stream(Writer::ROOT, "alpha", vec![]).is_err());
+    assert!(w.stream(Writer::ROOT, "Alph", vec![]).is_ok());
+    assert!(w.stream(Writer::ROOT, "Alpha2", vec![]).is_ok());
+    assert!(w.stream(a, "Alpha", vec![]).is_ok(), "same name in another storage");
+    assert!(w.stream(a, "Alpha", vec![]).is_err());
+    w.stream(Writer::ROOT, "Ünï", vec![]).unwrap();
+    assert!(w.stream(Writer::ROOT, "ÜNÏ", vec![]).is_err(), "non-ASCII case");
+}