diff --git a/lib/src/conflicts.rs b/lib/src/conflicts.rs
index 1ffd73ea1d..3d3bcf8fc7 100644
--- a/lib/src/conflicts.rs
+++ b/lib/src/conflicts.rs
@@ -14,9 +14,11 @@
 
 #![allow(missing_docs)]
 
+use std::borrow::Cow;
 use std::io::{Read, Write};
 use std::iter::zip;
 
+use bstr::ByteVec;
 use futures::{StreamExt, TryStreamExt};
 use itertools::Itertools;
 use regex::bytes::{Regex, RegexBuilder};
@@ -51,6 +53,12 @@ static CONFLICT_MARKER_REGEX: once_cell::sync::Lazy<Regex> = once_cell::sync::La
         .build()
         .unwrap()
 });
+static DIRECTIVE_LINE_REGEX: once_cell::sync::Lazy<Regex> = once_cell::sync::Lazy::new(|| {
+    RegexBuilder::new(r"^\\JJ")
+        .multi_line(true)
+        .build()
+        .unwrap()
+});
 
 fn write_diff_hunks(hunks: &[DiffHunk], file: &mut dyn Write) -> std::io::Result<()> {
     for hunk in hunks {
@@ -217,6 +225,12 @@ pub fn materialize_merge_result(
     single_hunk: Merge<ContentHunk>,
     output: &mut dyn Write,
 ) -> std::io::Result<()> {
+    let single_hunk: Merge<ContentHunk> =
+        single_hunk.into_map(|ContentHunk(side)| ContentHunk(encode_unmaterializeable_lines(side)));
+
+    // From now on, we assume some invariants guaranteed by the encoding function
+    // above. See its docstring for details.
+
     let slices = single_hunk.map(|content| content.0.as_slice());
     let merge_result = files::merge(&slices);
     match merge_result {
@@ -347,7 +361,11 @@ pub fn parse_merge_result(input: &[u8], num_sides: usize) -> Option<Merge<Conten
         }
     }
 
-    Some(result.into_map(ContentHunk))
+    Some(
+        result
+            .into_map(decode_materialized_side)
+            .into_map(ContentHunk),
+    )
 }
 
 fn parse_conflict_into_list_of_hunks(
@@ -540,3 +558,104 @@ pub async fn update_from_content(
     };
     Ok(new_file_ids)
 }
+
+const JJ_NO_NEWLINE_AT_EOF: &[u8] = b"\n\\JJ: No newline at the end of file\n";
+const JJ_VERBATIM_LINE: &str = "\\JJ Verbatim Line:";
+static JJ_VERBATIM_LINE_REPLACEMENT: once_cell::sync::Lazy<Vec<u8>> =
+    once_cell::sync::Lazy::new(|| format!("{JJ_VERBATIM_LINE}$0").into_bytes());
+static VERBATIM_LINE_REGEX: once_cell::sync::Lazy<Regex> = once_cell::sync::Lazy::new(|| {
+    RegexBuilder::new(&format!("^{}", regex::escape(JJ_VERBATIM_LINE)))
+        .multi_line(true)
+        .build()
+        .unwrap()
+});
+
+/// Encode a side of a conflict to satisfy some invariants
+///
+/// - The result will not contain any conflict markers
+/// - The result will contain 0 or more newline-terminated lines. In other
+///   words, it is either empty or ends in a newline.
+///
+/// This transformation is reversible and it is hoped that the result is
+/// human-readable. See the tests below for examples.
+fn encode_unmaterializeable_lines(mut side: Vec<u8>) -> Vec<u8> {
+    // TODO(ilyagr): It's likely the compiler won't be able to avoid a clone in
+    // the no-replacements case. If we care about that, we could try one of the
+    // approaches discussed in https://github.com/rust-lang/regex/issues/676.
+    // This also applies to the decoding function.
+    side = DIRECTIVE_LINE_REGEX
+        .replace_all(&Cow::from(side), JJ_VERBATIM_LINE_REPLACEMENT.as_slice())
+        .to_vec();
+    side = CONFLICT_MARKER_REGEX
+        .replace_all(&Cow::from(side), JJ_VERBATIM_LINE_REPLACEMENT.as_slice())
+        .to_vec();
+    if !side.is_empty() && !side.ends_with(b"\n") {
+        side.push_str(JJ_NO_NEWLINE_AT_EOF);
+    }
+    side
+}
+
+/// Undo the transformation done by `encode_unmaterializeable_lines`
+fn decode_materialized_side(mut side: Vec<u8>) -> Vec<u8> {
+    if side.ends_with(JJ_NO_NEWLINE_AT_EOF) {
+        side.truncate(side.len() - JJ_NO_NEWLINE_AT_EOF.len());
+    }
+    side = VERBATIM_LINE_REGEX
+        .replace_all(&Cow::from(side), b"")
+        .to_vec();
+    side
+}
+
+#[cfg(test)]
+mod test {
+    use bstr::BString;
+    use indoc::indoc;
+
+    use super::{decode_materialized_side, encode_unmaterializeable_lines};
+
+    #[test]
+    fn test_conflict_side_encoding_and_decoding() {
+        let initial_text: BString = indoc! {br"
+        <<<<<<<
+        blahblah"}
+        .into();
+
+        let encoded_text: BString = encode_unmaterializeable_lines(initial_text.to_vec()).into();
+        insta::assert_snapshot!(encoded_text, @r###"
+        \JJ Verbatim Line:<<<<<<<
+        blahblah
+        \JJ: No newline at the end of file
+        "###);
+        assert_eq!(
+            BString::from(decode_materialized_side(encoded_text.clone().into())),
+            initial_text
+        );
+
+        let doubly_encoded_text: BString =
+            encode_unmaterializeable_lines(encoded_text.clone().into()).into();
+        insta::assert_snapshot!(doubly_encoded_text, @r###"
+        \JJ Verbatim Line:\JJ Verbatim Line:<<<<<<<
+        blahblah
+        \JJ Verbatim Line:\JJ: No newline at the end of file
+        "###);
+        assert_eq!(
+            BString::from(decode_materialized_side(doubly_encoded_text.into())),
+            encoded_text
+        );
+    }
+
+    #[test]
+    fn test_conflict_side_encoding_and_decoding2() {
+        let initial_text = br"\JJ: No newline at the end of file";
+
+        let encoded_text: BString = encode_unmaterializeable_lines(initial_text.to_vec()).into();
+        insta::assert_snapshot!(encoded_text, @r###"
+        \JJ Verbatim Line:\JJ: No newline at the end of file
+        \JJ: No newline at the end of file
+        "###);
+        assert_eq!(
+            BString::from(decode_materialized_side(encoded_text.clone().into())),
+            BString::from(initial_text)
+        );
+    }
+}
diff --git a/lib/tests/test_conflicts.rs b/lib/tests/test_conflicts.rs
index 79325b1e10..bd8ff21789 100644
--- a/lib/tests/test_conflicts.rs
+++ b/lib/tests/test_conflicts.rs
@@ -326,6 +326,96 @@ fn test_materialize_parse_roundtrip() {
     "###);
 }
 
+#[test]
+fn test_materialize_parse_roundtrip_tricky() {
+    let test_repo = TestRepo::init();
+    let store = test_repo.repo.store();
+
+    let path = RepoPath::from_internal_string("file");
+    let base_id = testutils::write_file(
+        store,
+        path,
+        indoc! {"
+            \\JJ Verbatim Line: fake verbatim line
+            line 1
+            line 2 <<<<<<<
+            <<<<<<< line 3
+            line 4
+            line 5"},
+    );
+    let left_id = testutils::write_file(
+        store,
+        path,
+        indoc! {"
+            \\JJ Verbatim Line: fake verbatim line
+            line 1 left
+            line 2 left
+            <<<<<<<< line 3
+            line 4
+            line 5 left"},
+    );
+    let right_id = testutils::write_file(
+        store,
+        path,
+        indoc! {"
+            \\JJ Verbatim Line: fake verbatim line
+            line 1 right
+            line 2
+            line 3
+            line 4 right
+            line 5 right
+        "},
+    );
+
+    let conflict = Merge::from_removes_adds(
+        vec![Some(base_id.clone())],
+        vec![Some(left_id.clone()), Some(right_id.clone())],
+    );
+    let materialized = materialize_conflict_string(store, path, &conflict);
+    insta::assert_snapshot!(
+        materialized,
+        @r###"
+    \JJ Verbatim Line:\JJ Verbatim Line: fake verbatim line
+    <<<<<<< Conflict 1 of 1
+    %%%%%%% Changes from base to side #1
+    -line 1
+    -line 2 <<<<<<<
+    -\JJ Verbatim Line:<<<<<<< line 3
+    +line 1 left
+    +line 2 left
+    +<<<<<<<< line 3
+     line 4
+    -line 5
+    +line 5 left
+     \JJ: No newline at the end of file
+    +++++++ Contents of side #2
+    line 1 right
+    line 2
+    line 3
+    line 4 right
+    line 5 right
+    >>>>>>> Conflict 1 of 1 ends
+    "###
+    );
+
+    // The first add should always be from the left side
+    insta::assert_debug_snapshot!(
+        parse_merge_result(materialized.as_bytes(), conflict.num_sides()),
+        @r###"
+    Some(
+        Conflicted(
+            [
+                "\\JJ Verbatim Line: fake verbatim line\nline 1 left\nline 2 left\n<<<<<<<< line 3\nline 4\nline 5 left",
+                "\\JJ Verbatim Line: fake verbatim line\nline 1\nline 2 <<<<<<<\n<<<<<<< line 3\nline 4\nline 5",
+                "\\JJ Verbatim Line: fake verbatim line\nline 1 right\nline 2\nline 3\nline 4 right\nline 5 right\n",
+            ],
+        ),
+    )
+    "###);
+
+    // TODO: update_from_conflict?
+}
+
 #[test]
 fn test_materialize_conflict_no_newlines_at_eof() {
     let test_repo = TestRepo::init();
@@ -344,16 +434,29 @@ fn test_materialize_conflict_no_newlines_at_eof() {
     insta::assert_snapshot!(materialized,
         @r###"
     <<<<<<< Conflict 1 of 1
-    %%%%%%% Changes from base to side #1
-    -base+++++++ Contents of side #2
-    right>>>>>>> Conflict 1 of 1 ends
+    +++++++ Contents of side #1
+    %%%%%%% Changes from base to side #2
+    -base
+    +right
+     \JJ: No newline at the end of file
+    >>>>>>> Conflict 1 of 1 ends
     "###
     );
-    // BUG(#3968): These conflict markers cannot be parsed
+    // These conflict markers can be parsed (issue #3968)
     insta::assert_debug_snapshot!(parse_merge_result(
         materialized.as_bytes(),
         conflict.num_sides()
-    ),@"None");
+    ),@r###"
+    Some(
+        Conflicted(
+            [
+                "",
+                "base",
+                "right",
+            ],
+        ),
+    )
+    "###);
 }
 
 #[test]