diff --git a/docs/BACKLOG.md b/docs/BACKLOG.md
index 7054057..7d0f0e7 100644
--- a/docs/BACKLOG.md
+++ b/docs/BACKLOG.md
@@ -38,3 +38,7 @@ Each of these has a design-spike doc under `docs/design/` — read the linked do
 - **Multi-Model Orchestration Workflow**: Fast first-pass model with escalation to a larger model for low-confidence chunks. See `docs/design/multi_model_orchestration.md`.
 - **Incremental Track-Changes Support**: Diff-only redaction of newly-added paragraphs in a revised document. See `docs/design/incremental_track_changes_redaction.md` — flags a real under-redaction risk that needs a mitigation plan before implementation.
 - **Automated "Redaction Rationale" Reporting**: LLM-generated plain-English explanations in the audit log. See `docs/design/redaction_rationale_reporting.md` (covers how to label LLM-generated rationale as inference, not verified fact).
+
+## 4. CI / Infrastructure
+
+- **Self-Hosted GitHub Actions Runner** ([PR #34](https://github.com/LegalMarc/Marcut/pull/34), open, deferred): Points `ci.yml`, `macos-build-verify.yml`, and `macos-full-e2e.yml` at a local `[self-hosted, marcut-local]` runner instead of GitHub-hosted `macos-14`. Deferred during the 2026-07-15 release-readiness pass: the nightly/tag E2E red streak that motivated looking at this again turned out to have a deterministic root cause unrelated to runner choice (a `//`-comment-stripping regex in `parse_llm_response()` corrupting any LLM-extracted entity containing a URL — fixed on `fix/json-comment-strip-url-corruption`), so switching runners isn't needed to restore green CI. Revisit only if GitHub-hosted runner cost/queue-time/resource limits become a concrete problem — the security tradeoff noted in the PR (self-hosted runners execute arbitrary code from workflow-triggering events on the host machine) means it shouldn't be merged just because it's sitting there.
diff --git a/src/python/marcut/model.py b/src/python/marcut/model.py
index 2a65fbc..35ab3fc 100644
--- a/src/python/marcut/model.py
+++ b/src/python/marcut/model.py
@@ -143,6 +143,50 @@ def _repair_unbalanced_json(json_str: str) -> Optional[str]:
     return repaired
 
 
+def _strip_line_comments_outside_strings(json_str: str) -> str:
+    """
+    Strip `//`-style line comments that appear outside JSON string literals.
+
+    A naive `re.sub(r'//.*$', '', ...)` would also truncate legitimate string
+    values that happen to contain `//` -- most commonly a URL entity like
+    `"https://legal.example"` extracted by the LLM -- corrupting otherwise-
+    valid JSON. This walks the string tracking whether we're inside a string
+    literal (respecting `\\"` escapes) and only treats `//` as a comment
+    start when outside one.
+    """
+    out: List[str] = []
+    in_string = False
+    escape = False
+    i = 0
+    n = len(json_str)
+    while i < n:
+        ch = json_str[i]
+        if in_string:
+            out.append(ch)
+            if escape:
+                escape = False
+            elif ch == '\\':
+                escape = True
+            elif ch == '"':
+                in_string = False
+            i += 1
+            continue
+        if ch == '"':
+            in_string = True
+            out.append(ch)
+            i += 1
+            continue
+        if ch == '/' and i + 1 < n and json_str[i + 1] == '/':
+            newline_pos = json_str.find('\n', i)
+            if newline_pos == -1:
+                break
+            i = newline_pos
+            continue
+        out.append(ch)
+        i += 1
+    return ''.join(out)
+
+
 def parse_llm_response(response_text: str) -> Dict[str, Any]:
     """
     Sanitize and parse an LLM response, raising json.JSONDecodeError when invalid.
@@ -166,7 +210,7 @@ def parse_llm_response(response_text: str) -> Dict[str, Any]:
         else:
             json_str = cleaned[start:end]
 
-    json_str = re.sub(r'(?m)//.*$', '', json_str)
+    json_str = _strip_line_comments_outside_strings(json_str)
     json_str = re.sub(r',\s*(\]|\})', r'\1', json_str)
 
     try:
diff --git a/tests/test_model.py b/tests/test_model.py
index 36ea155..38c2ca4 100644
--- a/tests/test_model.py
+++ b/tests/test_model.py
@@ -84,9 +84,38 @@ def test_parse_json_with_comments(self):
             {"text": "John", "type": "NAME"} // This is a name
         ]}'''
         result = parse_llm_response(response)
-        
+
         assert 'entities' in result
-    
+
+    def test_parse_json_with_url_entity_containing_double_slash(self):
+        """A `//`-containing string value (e.g. a URL entity) must survive
+        comment-stripping intact -- a naive `//.*$` regex would truncate the
+        string mid-value and corrupt the JSON (root cause of the nightly E2E
+        failure streak: the LLM extracting "https://legal.example" as an
+        entity)."""
+        response = json.dumps({
+            "entities": [
+                {"text": "alice@example.com", "type": "NAME"},
+                {"text": "https://legal.example", "type": "ORG"},
+            ]
+        })
+        result = parse_llm_response(response)
+
+        assert len(result['entities']) == 2
+        assert result['entities'][1]['text'] == 'https://legal.example'
+
+    def test_parse_json_with_comment_after_url_entity(self):
+        """A genuine trailing `//` comment after a URL-containing entity line
+        must still be stripped, without corrupting the URL itself."""
+        response = '''{"entities": [
+            {"text": "https://legal.example", "type": "ORG"}, // a URL
+            {"text": "Jane Doe", "type": "NAME"}
+        ]}'''
+        result = parse_llm_response(response)
+
+        assert result['entities'][0]['text'] == 'https://legal.example'
+        assert result['entities'][1]['text'] == 'Jane Doe'
+
     def test_parse_empty_entities(self):
         """Test parsing response with no entities."""
         response = '{"entities": []}'