diff --git a/CHANGELOG.md b/CHANGELOG.md
index cd3dd6c1..5a7c8868 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -106,6 +106,20 @@ All notable changes to vouch are documented here. Format follows
   artifact the caller could not already retrieve, and it touches no write path.
 
 ### Fixed
+- **`kb.experts` no longer leaks out-of-scope claims into entity rankings**
+  (#714): `rank_experts` aggregated evidence density over every claim in
+  the KB with no viewer/scope filtering at all, unlike every sibling
+  claim-aggregating read surface (`context.py`, `graph.py`, `digest.py`,
+  `health.py`, `compile.py`, and `themes.detect_themes`, the closest
+  shape-wise sibling). A `project`- or `agent`-scoped claim the caller
+  cannot otherwise retrieve still inflated `claim_count`, `citation_count`,
+  and `score`, and could surface verbatim in `top_claim_ids` — handing the
+  caller a claim id it cannot fetch. `rank_experts` now takes an optional
+  `viewer` (defaulting to `scoping.viewer_from(...)`, matching
+  `detect_themes`) and filters through `scoping.is_visible` before a claim
+  can contribute anything, with the FTS candidate fetch run through
+  `scoping.scoped_fetch_limit` so a mostly-out-of-scope KB doesn't starve
+  the candidate pool before the filter runs.
 - **`vouch render-wiki` drops archived pages** (#695):
   `render_wiki_cmd` passed every on-disk page into index/MOC, so retired
   titles kept wiki links after archive. the CLI now filters to the same
diff --git a/src/vouch/experts.py b/src/vouch/experts.py
index c20f409e..00ac6a26 100644
--- a/src/vouch/experts.py
+++ b/src/vouch/experts.py
@@ -15,6 +15,7 @@
 from . import index_db
 from .models import Claim, ClaimStatus, utcnow
 from .salience import _substring_entity_ids
+from .scoping import ViewerContext, is_visible, scoped_fetch_limit, viewer_from
 from .storage import KBStore
 
 # A superseded / archived / redacted claim is not live evidence and must never
@@ -44,24 +45,37 @@ def rank_experts(
     limit: int = 10,
     min_claims: int = 1,
     weight: str = "count",
+    viewer: ViewerContext | None = None,
 ) -> list[dict[str, Any]]:
     """Return entities ranked by evidence density on ``topic``.
 
     ``weight`` is one of ``count`` | ``recency`` | ``citation``; an unknown
     value falls back to ``count`` (never raises), matching the defensive-config
     style used elsewhere. Ordered by descending score with a stable tie-break
-    on ``entity_id``.
+    on ``entity_id``. ``viewer`` defaults to the config-resolved context, so a
+    KB read with no explicit viewer reads as its own project (see
+    ``scoping.viewer_from``) — matching ``themes.detect_themes``, the sibling
+    read surface this mirrors. Claims the viewer cannot retrieve never
+    contribute to ``claim_count``, ``citation_count``, or ``score``, and never
+    appear in ``top_claim_ids``: scoring after filtering, not just scrubbing
+    the id list, is what keeps a mostly-private entity from outranking one
+    the viewer can actually read.
     """
     if weight not in _VALID_WEIGHTS:
         weight = "count"
+    if viewer is None:
+        viewer = viewer_from(config_path=store.config_path)
 
     entities = store.list_entities()
     by_id = {ent.id: ent for ent in entities}
     topic_entity_ids = set(_substring_entity_ids(entities, topic))
 
     # Candidate claims: FTS hits on the topic, plus every claim that references
-    # an entity whose name/alias matches the topic.
-    fetch = max(limit * 5, 50)
+    # an entity whose name/alias matches the topic. Over-fetched via
+    # scoped_fetch_limit so a viewer-scoped KB whose top FTS hits are mostly
+    # private doesn't starve the candidate pool before the scope filter below
+    # ever runs.
+    fetch = scoped_fetch_limit(max(limit * 5, 50), viewer)
     fts_claim_ids = {
         cid
         for kind, cid, _snip, _score in index_db.search(store.kb_dir, topic, limit=fetch)
@@ -77,6 +91,8 @@ def rank_experts(
     for claim in store.list_claims():
         if claim.status in _EXCLUDED_STATUSES:
             continue
+        if not is_visible(claim.scope, viewer):
+            continue
         matched = claim.id in fts_claim_ids or bool(
             set(claim.entities) & topic_entity_ids
         )
diff --git a/tests/test_experts.py b/tests/test_experts.py
index 0e3536e2..f17366ac 100644
--- a/tests/test_experts.py
+++ b/tests/test_experts.py
@@ -14,7 +14,8 @@
 
 from vouch.experts import rank_experts
 from vouch.jsonl_server import handle_request
-from vouch.models import Claim, ClaimStatus, Entity, EntityType
+from vouch.models import ArtifactScope, Claim, ClaimStatus, Entity, EntityType, Visibility
+from vouch.scoping import ViewerContext
 from vouch.storage import KBStore
 
 
@@ -127,6 +128,51 @@ def test_deterministic_tie_break_on_entity_id(store: KBStore) -> None:
     assert tied == ["a1", "a2"]  # equal score -> ascending entity_id
 
 
+def test_scopes_ranking_to_the_viewer(store: KBStore) -> None:
+    """A viewer outside a claim's project must not have that claim inflate
+    claim_count/citation_count/score, and must never see its id surface in
+    top_claim_ids — scoring after filtering, not just scrubbing the id list,
+    is what keeps a mostly-private entity from outranking one the viewer can
+    actually read."""
+    src = store.put_source(b"y")
+    store.put_entity(Entity(id="jwt", name="JWT", type=EntityType.CONCEPT))
+    store.put_entity(Entity(id="alice", name="alice", type=EntityType.PERSON))
+    for i in range(3):
+        store.put_claim(
+            Claim(
+                id=f"priv{i}",
+                text=f"jwt fact {i} by alice",
+                evidence=[src.id],
+                entities=["jwt", "alice"],
+                scope=ArtifactScope(visibility=Visibility.PROJECT, project="secret-project"),
+            )
+        )
+    store.put_claim(
+        Claim(
+            id="pub",
+            text="jwt public fact by alice",
+            evidence=[src.id],
+            entities=["jwt", "alice"],
+            scope=ArtifactScope(visibility=Visibility.PUBLIC),
+        )
+    )
+
+    foreign = ViewerContext(project="other-project", agent=None)
+    rows = rank_experts(store, "JWT", viewer=foreign)
+    alice = next(r for r in rows if r["entity_id"] == "alice")
+    assert alice["claim_count"] == 1
+    assert alice["top_claim_ids"] == ["pub"]
+    for row in rows:
+        assert "priv0" not in row["top_claim_ids"]
+        assert "priv1" not in row["top_claim_ids"]
+        assert "priv2" not in row["top_claim_ids"]
+
+    owner = ViewerContext(project="secret-project", agent=None)
+    owner_rows = rank_experts(store, "JWT", viewer=owner)
+    owner_alice = next(r for r in owner_rows if r["entity_id"] == "alice")
+    assert owner_alice["claim_count"] == 4
+
+
 def test_jsonl_experts_envelope_success(store: KBStore, monkeypatch) -> None:
     # kb.experts over the JSONL contract: a well-formed request returns the
     # {id, ok, result} envelope with the ranking under result["experts"].