diff --git a/src/benchmark_radar/authors.py b/src/benchmark_radar/authors.py
index f48a372a..d86b6ce9 100644
--- a/src/benchmark_radar/authors.py
+++ b/src/benchmark_radar/authors.py
@@ -132,13 +132,31 @@ def popular_repositories(
     a week is a real signal; so is a highly starred one seen once.
     """
     corpus = build_corpus(snapshots)
+    observed_repos: dict[str, dict[str, str]] = {}
+    for observation in corpus["observations"]:
+        match = _GITHUB_REPO.match(str(observation.get("url") or ""))
+        if match:
+            owner, name = match.group(1), match.group(2).removesuffix(".git")
+            full_name = f"{owner}/{name}"
+            observed_repos.setdefault(observation["entity_id"], {})[full_name.casefold()] = (
+                f"https://github.com/{full_name}"
+            )
     ranked = []
     for entity in corpus["entities"]:
         if entity["type"] != "artifact":
             continue
-        match = _GITHUB_REPO.match(str(entity.get("url") or ""))
+        repo_url = str(entity.get("url") or "")
+        match = _GITHUB_REPO.match(repo_url)
         if not match:
-            continue
+            candidates = observed_repos.get(entity["id"], {})
+            # Recover only an actually observed, unambiguous repository. A
+            # paper's links alone must not invent seeds or assign stars to one
+            # of several distinct repositories joined by the artifact graph.
+            if len(candidates) != 1:
+                continue
+            repo_url = next(iter(candidates.values()))
+            match = _GITHUB_REPO.match(repo_url)
+            assert match is not None
         owner, name = match.group(1), match.group(2).removesuffix(".git")
         if owner.lower() in {"apps", "marketplace", "sponsors", "topics"}:
             continue
@@ -155,7 +173,7 @@ def popular_repositories(
                 "full_name": f"{owner}/{name}",
                 "owner": owner,
                 "name": name,
-                "url": entity.get("url"),
+                "url": repo_url,
                 "title": entity.get("label"),
                 "stars": float(metrics.get("stars") or 0),
                 "seen_days": len(entity.get("seen_days") or []),
diff --git a/tests/test_authors.py b/tests/test_authors.py
index 7bb15d5e..756db7af 100644
--- a/tests/test_authors.py
+++ b/tests/test_authors.py
@@ -163,3 +163,50 @@ def test_contacts_csv_contains_every_contact_and_flattens_lists():
     assert "data quality; dataset" in rendered
     assert "org/one; org/two" in rendered
     assert "other" in rendered
+
+
+def test_survey_keeps_observed_repository_when_corpus_prefers_linked_paper():
+    repo = _repo("example/benchmark", stars=100, categories=["benchmark"])
+    paper = RadarItem(
+        source="arXiv",
+        source_id="2608.12345",
+        title="A benchmark paper",
+        url="https://arxiv.org/abs/2608.12345",
+        published_at=repo.published_at,
+        categories=["benchmark"],
+        artifact_urls=[repo.url],
+        summary="A scored benchmark with an observed repository.",
+    )
+    current = snapshot_for_run(_run([repo, paper]))
+    ranked = authors.popular_repositories([current])
+    assert [row["full_name"] for row in ranked] == ["example/benchmark"]
+    assert ranked[0]["stars"] == 100
+    assert ranked[0]["url"] == repo.url
+
+
+def test_survey_does_not_treat_unobserved_paper_links_as_repository_seeds():
+    paper = RadarItem(
+        source="arXiv",
+        source_id="2608.12345",
+        title="A benchmark paper",
+        url="https://arxiv.org/abs/2608.12345",
+        published_at=datetime(2026, 8, 4, tzinfo=UTC),
+        categories=["benchmark"],
+        artifact_urls=["https://github.com/example/unobserved"],
+    )
+    assert authors.popular_repositories([snapshot_for_run(_run([paper]))]) == []
+
+
+def test_survey_does_not_assign_aggregate_stars_to_ambiguous_repository_aliases():
+    first = _repo("one/benchmark", stars=100, categories=["benchmark"])
+    second = _repo("two/benchmark", stars=50, categories=["benchmark"])
+    paper = RadarItem(
+        source="arXiv",
+        source_id="2608.12345",
+        title="A benchmark paper",
+        url="https://arxiv.org/abs/2608.12345",
+        published_at=first.published_at,
+        categories=["benchmark"],
+        artifact_urls=[first.url, second.url],
+    )
+    assert authors.popular_repositories([snapshot_for_run(_run([first, second, paper]))]) == []