diff --git a/src/uu/sort/src/chunks.rs b/src/uu/sort/src/chunks.rs
index 3387817b06a..ca9efcd13e8 100644
--- a/src/uu/sort/src/chunks.rs
+++ b/src/uu/sort/src/chunks.rs
@@ -20,9 +20,7 @@ use self_cell::self_cell;
 use uucore::error::{UResult, USimpleError, strip_errno};
 use uucore::translate;
 
-use crate::{
-    GeneralBigDecimalParseResult, GlobalSettings, Line, SortMode, numeric_str_cmp::NumInfo,
-};
+use crate::{GeneralBigDecimalParseResult, GlobalSettings, Line, numeric_str_cmp::NumInfo};
 
 const ALLOC_CHUNK_SIZE: usize = 64 * 1024;
 const MAX_TOKEN_BUFFER_BYTES: usize = 4 * 1024 * 1024;
@@ -312,7 +310,7 @@ fn parse_lines<'a>(
             .parsed_floats
             .reserve(estimated.saturating_mul(settings.precomputed.floats_per_line));
     }
-    if settings.mode == SortMode::Numeric {
+    if settings.precomputed.whole_line_numeric {
         line_data.line_num_floats.reserve(estimated);
     }
     let mut start = 0usize;
diff --git a/src/uu/sort/src/sort.rs b/src/uu/sort/src/sort.rs
index 7a8bd550fdb..f47f457eda9 100644
--- a/src/uu/sort/src/sort.rs
+++ b/src/uu/sort/src/sort.rs
@@ -332,6 +332,7 @@ struct Precomputed {
     fast_lexicographic: bool,
     fast_locale_collation: bool,
     fast_ascii_insensitive: bool,
+    whole_line_numeric: bool,
     tokenize_blank_thousands_sep: bool,
     tokenize_allow_unit_after_blank: bool,
 }
@@ -397,6 +398,27 @@ impl GlobalSettings {
         self.precomputed.fast_locale_collation =
             disable_fast_lexicographic && self.can_use_fast_lexicographic();
         self.precomputed.fast_ascii_insensitive = self.can_use_fast_ascii_insensitive();
+        self.precomputed.whole_line_numeric = self.can_use_whole_line_numeric();
+    }
+
+    /// Returns true when a number parsed from the whole line can stand in for
+    /// the key.
+    ///
+    /// `-n` parses the line once up front and compares those numbers before
+    /// looking at any key. That is only the same comparison when the single
+    /// key spans the entire line: `-n -k1.2` sorts on the line's second
+    /// character onwards, and `-n -t. -k2` on its second field, neither of
+    /// which the line as a whole stands for. A key of its own `r` also has to
+    /// go the long way, since the shortcut only knows the global one.
+    fn can_use_whole_line_numeric(&self) -> bool {
+        self.mode == SortMode::Numeric && self.selectors.len() == 1 && {
+            let selector = &self.selectors[0];
+            selector.settings.mode == SortMode::Numeric
+                && !selector.settings.reverse
+                && selector.from.field == 1
+                && selector.from.char == 1
+                && selector.to.is_none()
+        }
     }
 
     /// Returns true when the fast lexicographic path can be used safely.
@@ -654,7 +676,7 @@ impl<'a> Line<'a> {
             || settings.precomputed.selections_per_line > 0
             || settings.precomputed.num_infos_per_line > 0
             || settings.precomputed.floats_per_line > 0
-            || settings.mode == SortMode::Numeric;
+            || settings.precomputed.whole_line_numeric;
         if !needs_line_data {
             return Self { line, index };
         }
@@ -667,7 +689,7 @@ impl<'a> Line<'a> {
                 &settings.precomputed,
             );
         }
-        if settings.mode == SortMode::Numeric {
+        if settings.precomputed.whole_line_numeric {
             // exclude inf, nan, scientific notation; GNU -n does not treat '+' as a sign
             let line_num_float = (!line.iter().any(u8::is_ascii_alphabetic))
                 .then(|| std::str::from_utf8(line).ok())
diff --git a/tests/by-util/test_sort.rs b/tests/by-util/test_sort.rs
index 0747ddc3935..44c59b90597 100644
--- a/tests/by-util/test_sort.rs
+++ b/tests/by-util/test_sort.rs
@@ -325,6 +325,56 @@ fn test_numeric_sort_rejects_leading_plus_sign() {
         .stdout_is("+1\n+10\n+2\n");
 }
 
+#[test]
+fn test_numeric_sort_uses_the_key_not_the_whole_line() {
+    // `-n` compares a number parsed from the line as a whole, which only
+    // stands for the key when the key is the whole line.
+
+    // Character offset: the key is "9" and "1", so 21 comes first. Comparing
+    // the lines instead puts 19 first.
+    new_ucmd!()
+        .args(&["-n", "-k1.2"])
+        .pipe_in("19\n21\n")
+        .succeeds()
+        .stdout_is("21\n19\n");
+
+    // Field: with '.' as the separator the key is "9" and "1" again, while
+    // both lines parse as numbers on their own.
+    new_ucmd!()
+        .args(&["-n", "-t.", "-k2"])
+        .pipe_in("1.9\n2.1\n")
+        .succeeds()
+        .stdout_is("2.1\n1.9\n");
+
+    // A key that reverses on its own: the shortcut only knows the global
+    // ordering, so it has to go the long way.
+    new_ucmd!()
+        .args(&["-n", "-k1r"])
+        .pipe_in("19\n21\n3\n")
+        .succeeds()
+        .stdout_is("3\n21\n19\n");
+
+    // Equal keys fall back to comparing the lines byte by byte, not
+    // numerically.
+    new_ucmd!()
+        .args(&["-n", "-k1.2"])
+        .pipe_in("1\n10\n2\n20\n3\n")
+        .succeeds()
+        .stdout_is("1\n10\n2\n20\n3\n");
+
+    // The whole-line cases keep working.
+    new_ucmd!()
+        .arg("-n")
+        .pipe_in("19\n21\n3\n")
+        .succeeds()
+        .stdout_is("3\n19\n21\n");
+    new_ucmd!()
+        .args(&["-n", "-k1"])
+        .pipe_in("19\n21\n3\n")
+        .succeeds()
+        .stdout_is("3\n19\n21\n");
+}
+
 #[test]
 fn test_check_zero_terminated_failure() {
     new_ucmd!()