diff --git a/crates/doc-binary/src/tests/word6.rs b/crates/doc-binary/src/tests/word6.rs
index 6bd9fb8f3..fa3f68a51 100644
--- a/crates/doc-binary/src/tests/word6.rs
+++ b/crates/doc-binary/src/tests/word6.rs
@@ -274,6 +274,19 @@ fn a_runs_font_charset_chooses_its_code_page() {
     assert_eq!(texts(&d), ["Da Да"]);
 }
 
+/// Text in an East Asian code page (a Japanese font, charset 128: Shift-JIS) is refused, not
+/// read a byte at a time into replacement characters (review of wordcraft#6345).
+#[test]
+fn east_asian_text_is_refused_not_lost() {
+    let fonts: &[(u8, &str)] = &[(0, "Times New Roman"), (128, "MS Mincho")];
+    let runs: &[(u32, u32, &[u8])] = &[(0, 2, &[0x5D, 0x01, 0x00]), (2, 3, &[])];
+    let d = Doc6 { text: b"\x93\xfa\r", fonts, runs, ..Default::default() };
+    assert!(matches!(read(&word6_file(&d)), Err(DocbinError::NotWord(e)) if e.contains("East Asian")));
+    // The same file in a Western font still opens.
+    let d = Doc6 { text: b"\x93\xfa\r", fonts, runs: &[], ..Default::default() };
+    assert_eq!(texts(&d).len(), 1);
+}
+
 #[test]
 fn the_font_sprm_is_found_after_other_sprms() {
     // As in real files: bold-like toggles, a size, a picture anchor's length-prefixed sprm.
@@ -302,15 +315,6 @@ fn an_unknown_sprm_leaves_font_0() {
     assert_eq!(texts(&d), ["Äà"]);
 }
 
-#[test]
-fn east_asian_font_bytes_read_as_replacement_characters() {
-    // A multibyte code page can't be read one byte per CP: ASCII stays, the rest is U+FFFD.
-    let text = b"ab\x82\xA0\r";
-    let runs: &[(u32, u32, &[u8])] = &[(0, 4, &[0x5D, 0x01, 0x00]), (4, 5, &[])];
-    let d = Doc6 { text, fonts: &[(0, "Arial"), (128, "MS Mincho")], runs, ..Default::default() };
-    assert_eq!(texts(&d), ["ab\u{FFFD}\u{FFFD}"]);
-}
-
 #[test]
 fn the_toggle_sprms_before_the_font_are_framed() {
     // 0x57 (strike, a one-byte toggle in [MS-DOC]'s Prm0 isprm table) before the font sprm: the
diff --git a/crates/doc-binary/src/word6.rs b/crates/doc-binary/src/word6.rs
index 0139e3ce5..e9fab2d74 100644
--- a/crates/doc-binary/src/word6.rs
+++ b/crates/doc-binary/src/word6.rs
@@ -31,8 +31,9 @@
 //!   others were identified against Word. See [`chp`] and [`pap`].
 //! - A font table entry is a size byte (the entry's length − 1), a flags byte, a 2-byte weight, the
 //!   charset byte, one more byte, then the name in 8-bit characters.
-//! - The stylesheet (pair 1) has the Word 97 shape with smaller parts: a 14-byte header whose
-//!   first two words are the style count and the size of a style's fixed part (8, Word 97: 10);
+//! - The stylesheet (pair 1) has the Word 97 shape with smaller parts: a u16 header size (14),
+//!   then a header of that many bytes whose first two words are the style count and the size of
+//!   a style's fixed part (8, Word 97: 10);
 //!   then per style a u16 size (0: an empty slot) and that many bytes: the fixed part (sti, then
 //!   kind and base style, then UPX count and next style, then a size), the name (a length byte,
 //!   8-bit characters, a 0), and the UPXs, each a u16 size and its bytes, padded to even. A
@@ -258,6 +259,13 @@ pub(crate) fn content(word: &[u8], fib: &Fib6) -> Result<Content, DocbinError> {
         _ => 0,
     };
     let code_pages = if windows { code_page_runs(&chp_runs, &pap_runs, &facts, &fonts, default_code_page) } else { Vec::new() };
+    // shortcut: East Asian (lead and trail byte) text is refused, as all Word 6.0/95 files were
+    // before; read it a character at a time when the piece walk can decode byte pairs.
+    let east_asian = || DocbinError::NotWord("Word 6.0/95 text in an East Asian code page is not read yet".into());
+    if craft_codepage::is_multibyte_code_page(default_code_page) || code_pages.iter().any(|r| craft_codepage::is_multibyte_code_page(r.code_page)) {
+        return Err(east_asian());
+    }
+    let multibyte_piece = std::cell::Cell::new(false);
     let text = Word6Text { wide: fib.wide, default_code_page, code_pages };
     let pieces = if fib.complex {
         let (fc, lcb) = fib.pair(PAIR_CLX).ok_or_else(|| DocbinError::Malformed("the fast-saved file has no piece table".into()))?;
@@ -267,12 +275,20 @@ pub(crate) fn content(word: &[u8], fib: &Fib6) -> Result<Content, DocbinError> {
             if !windows || prm & 1 == 0 {
                 return None;
             }
-            last_font(prcs.get(usize::from(prm >> 1))?).and_then(|f| font_code_page(&fonts, f))
+            let cp = last_font(prcs.get(usize::from(prm >> 1))?).and_then(|f| font_code_page(&fonts, f));
+            if cp.is_some_and(craft_codepage::is_multibyte_code_page) {
+                multibyte_piece.set(true);
+            }
+            cp
         };
         // The Prcs then carry Word 97 sprms, as the CHPXs and PAPXs do: the walk applies their
         // character sprms to the piece's text and their paragraph sprms to the paragraph whose mark
         // the piece holds.
-        PieceTable::parse_word6(word, fc, lcb, text, piece_code_page, translate_prc)?
+        let table = PieceTable::parse_word6(word, fc, lcb, text, piece_code_page, translate_prc)?;
+        if multibyte_piece.get() {
+            return Err(east_asian());
+        }
+        table
     } else {
         // Word reads no more characters than the text's byte range (fcMin..fcMac) holds, one byte
         // each or two for UTF-16, whatever the character count says.