diff --git a/crates/text/src/layout.rs b/crates/text/src/layout.rs
index 26ae3d33..f844b571 100644
--- a/crates/text/src/layout.rs
+++ b/crates/text/src/layout.rs
@@ -1035,6 +1035,18 @@ fn is_emoji(c: char) -> bool {
     matches!(c, '\u{1F1E6}'..='\u{1F1FF}' | '\u{1F300}'..='\u{1FAFF}')
 }
 
+/// The lower case of the capital sigma at `i`: the final form ς or σ, chosen from the letters
+/// around it in the whole text (as lower-casing the whole string would), not from the lone char.
+fn lower_sigma(chars: &[char], i: usize) -> char {
+    // Enough context to step over any realistic run of combining marks next to the sigma.
+    const CONTEXT: usize = 32;
+    let before: String = chars.get(i.saturating_sub(CONTEXT)..i).unwrap_or(&[]).iter().collect();
+    let after: String = chars.iter().skip(i + 1).take(CONTEXT).collect();
+    let lower = format!("{before}\u{3A3}{after}").to_lowercase();
+    // Only the sigma's lower case depends on context, so the text before it keeps its length.
+    lower.get(before.to_lowercase().len()..).and_then(|s| s.chars().next()).unwrap_or('σ')
+}
+
 /// One run of chars shaped with a single face and style.
 struct Segment<'a> {
     start: usize,
@@ -1065,6 +1077,8 @@ fn shape_segment(chars: &[char], seg: &Segment, face: &Arc<Face>, features: &[Fe
             for u in c.to_uppercase() {
                 mapped.push((u, src, true));
             }
+        } else if style.case == TextCase::Lower && *c == 'Σ' {
+            mapped.push((lower_sigma(chars, src), src, false));
         } else {
             let mut buf = [0u8; 4];
             for u in style.case.apply_after(c.encode_utf8(&mut buf), prev).chars() {
diff --git a/crates/text/src/tests.rs b/crates/text/src/tests.rs
index ebf5889c..4475fbbc 100644
--- a/crates/text/src/tests.rs
+++ b/crates/text/src/tests.rs
@@ -497,3 +497,25 @@ fn the_caret_after_a_line_break_sits_on_the_next_line() {
     let l = layout(&TextData::new("ab\n"), None, None);
     assert_eq!(l.caret(3), (0.0, 1));
 }
+
+#[test]
+fn lower_case_writes_final_sigma() {
+    if !have_fonts() {
+        return;
+    }
+    let gids = |t: &TextData| layout(t, None, None).glyphs.iter().map(|g| (g.gid, g.char_start)).collect::<Vec<_>>();
+    let want = gids(&TextData::new("οδος ας"));
+    let Some(face) = layout(&TextData::new("ς"), None, None).glyphs.first().map(|g| g.face.clone()) else { return };
+    if face.glyph_for('σ') == face.glyph_for('ς') {
+        eprintln!("face has no distinct final sigma; skipping");
+        return;
+    }
+    let mut t = TextData::new("ΟΔΟΣ ΑΣ");
+    t.style.case = figcraft_doc::TextCase::Lower;
+    assert_eq!(gids(&t), want, "one run");
+    // The word split across two styled runs: the final form is still chosen from the whole text.
+    let mut other = t.style.clone();
+    other.font_size += 4.0;
+    t.runs.push(figcraft_doc::StyleRun { start: 3, end: 4, style: other });
+    assert_eq!(gids(&t).iter().map(|g| g.0).collect::<Vec<_>>(), want.iter().map(|g| g.0).collect::<Vec<_>>(), "split runs");
+}