diff --git a/eval/dataset/REVIEWBOT_EVAL.md b/eval/dataset/REVIEWBOT_EVAL.md
index 95720588..3c4d1f5a 100644
--- a/eval/dataset/REVIEWBOT_EVAL.md
+++ b/eval/dataset/REVIEWBOT_EVAL.md
@@ -117,7 +117,9 @@ interpreter or manifest and does not probe installed packages.
   `changed_files` == the frozen GT diff's file count.
 - Sanitized artifacts (markers stripped, neutral heading) must fit the
   judge's 24k cap — over-cap is a campaign failure, not a silent
-  truncation.
+  truncation. The archived adapter captures the actual inline comment payload
+  from the ReviewBot publisher and appends it to the evaluation artifact;
+  ReviewBot's ordinary review-body artifact omits placed inline findings.
 - Before reporting, `build_reviewbot_report.py --verify` audits exact
   denominators, pinned judge identity, and fresh candidate sha256s in
   every verdict (judge_val skips existing files, so its exit code is not
diff --git a/eval/dataset/reviewbot_archived_replay.py b/eval/dataset/reviewbot_archived_replay.py
index 42cd5959..7d075190 100644
--- a/eval/dataset/reviewbot_archived_replay.py
+++ b/eval/dataset/reviewbot_archived_replay.py
@@ -81,6 +81,48 @@ def historical_pull(pull, *, pr: int, head: str, base: str):
     )
 
 
+def shadow_candidate(body: str, inline_comments: list[dict]) -> str:
+    """Include the inline payload that a COMMENT review would publish.
+
+    The normal review-body artifact intentionally omits placed findings. A
+    judge reading that artifact alone would score a review with inline
+    findings as though it made none.
+    """
+    if not inline_comments:
+        return body
+    if "No actionable findings." in body:
+        raise ValueError("review body contradicts its inline findings")
+    parts = [body.rstrip(), "", "### Inline findings", ""]
+    for comment in inline_comments:
+        parts.extend([
+            f"#### {comment['path']}:{comment['line']} ({comment['side']})",
+            "",
+            str(comment["body"]).strip(),
+            "",
+        ])
+    return "\n".join(parts).rstrip() + "\n"
+
+
+def write_shadow_artifact(outcome, inline_comments: list[dict] | None, state_dir: Path) -> None:
+    """Complete only this evaluation's shadow artifact with placed comments."""
+    if outcome.status != "shadow":
+        return
+    metrics = outcome.metrics or {}
+    if inline_comments is None or metrics.get("inline_placed") != len(inline_comments):
+        raise RuntimeError("shadow inline payload differs from publication metrics")
+    if "See inline comments below." in outcome.body and not inline_comments:
+        raise RuntimeError("review body references missing inline findings")
+    if outcome.artifact_path is None:
+        raise RuntimeError("shadow review has no artifact")
+    artifact = Path(outcome.artifact_path).resolve(strict=True)
+    if artifact.parent != (state_dir / "artifacts").resolve():
+        raise RuntimeError("shadow artifact escaped the evaluation state")
+    artifact.write_text(
+        shadow_candidate(artifact.read_text(encoding="utf-8"), inline_comments),
+        encoding="utf-8",
+    )
+
+
 def main() -> int:
     parser = argparse.ArgumentParser(description=__doc__)
     parser.add_argument("--pr", type=int, required=True)
@@ -103,6 +145,7 @@ def main() -> int:
     from omni_reviewbot.cli import build_runtime
     from omni_reviewbot.config import Settings
     from omni_reviewbot.models import ChangedFile
+    from omni_reviewbot import publication
 
     settings = Settings.from_env()
     if settings.post or settings.review_context_mode != "no_discussion":
@@ -138,9 +181,23 @@ def list_files(number):
 
     github.get_pull = get_pull
     github.list_files = list_files
-    outcome = runtime.pipeline.run(
-        pr_number=args.pr, expected_head=args.expected_head, post=False,
-    )
+    original_inline = publication.inline_review_comments
+    captured: list[dict] | None = None
+
+    def capture_inline(*call_args, **call_kwargs):
+        nonlocal captured
+        comments, indexes = original_inline(*call_args, **call_kwargs)
+        captured = list(comments)
+        return comments, indexes
+
+    publication.inline_review_comments = capture_inline
+    try:
+        outcome = runtime.pipeline.run(
+            pr_number=args.pr, expected_head=args.expected_head, post=False,
+        )
+    finally:
+        publication.inline_review_comments = original_inline
+    write_shadow_artifact(outcome, captured, settings.state_dir)
     print(outcome.body)
     print(f"status: {outcome.status}")
     metrics = outcome.metrics or {}
diff --git a/test/test_reviewbot_eval.py b/test/test_reviewbot_eval.py
index 7ddcbfca..08b72b8b 100644
--- a/test/test_reviewbot_eval.py
+++ b/test/test_reviewbot_eval.py
@@ -11,6 +11,7 @@
 import zipfile
 from dataclasses import replace
 from pathlib import Path
+from types import SimpleNamespace
 
 import pytest
 
@@ -100,6 +101,48 @@ class Pull:
             archive.historical_pull(stale, pr=4893, head="a" * 40, base="d" * 40)
 
 
+def test_shadow_candidate_includes_the_posted_inline_payload():
+    body = "## Omni ReviewBot review\n\nSummary\n\nSee inline comments below.\n"
+    comment = {
+        "path": "models/foo.py", "line": 42, "side": "RIGHT",
+        "body": "**[major] Missing guard**\n\nThis path can fail.",
+    }
+    rendered = archive.shadow_candidate(body, [comment])
+    assert "### Inline findings" in rendered
+    assert "models/foo.py:42 (RIGHT)" in rendered
+    assert comment["body"] in rendered
+    assert archive.shadow_candidate(body, []) == body
+    with pytest.raises(ValueError, match="contradicts"):
+        archive.shadow_candidate("No actionable findings.\n", [comment])
+
+
+def test_shadow_artifact_requires_publication_parity_and_evaluation_path(tmp_path):
+    state = tmp_path / "state"
+    artifacts = state / "artifacts"
+    artifacts.mkdir(parents=True)
+    artifact = artifacts / "pr-4810-head.md"
+    artifact.write_text("See inline comments below.\n")
+    comment = {
+        "path": "foo.py", "line": 7, "side": "RIGHT",
+        "body": "**[P2] Missing guard**\n\nReason",
+    }
+    outcome = SimpleNamespace(
+        status="shadow", metrics={"inline_placed": 1},
+        body="See inline comments below.\n", artifact_path=artifact,
+    )
+    with pytest.raises(RuntimeError, match="differs"):
+        archive.write_shadow_artifact(outcome, [], state)
+    assert artifact.read_text() == outcome.body
+    outcome.artifact_path = tmp_path / "outside.md"
+    outcome.artifact_path.write_text(outcome.body)
+    with pytest.raises(RuntimeError, match="escaped"):
+        archive.write_shadow_artifact(outcome, [comment], state)
+    assert outcome.artifact_path.read_text() == outcome.body
+    outcome.artifact_path = artifact
+    archive.write_shadow_artifact(outcome, [comment], state)
+    assert comment["body"] in artifact.read_text()
+
+
 def test_archived_runner_identity_and_command_are_explicit():
     release = {"paired": True}
     config = arm.build_config(