diff --git a/services/_shared/personas/code_reviewer/dispatch.py b/services/_shared/personas/code_reviewer/dispatch.py
index 32e14599..8efa40ff 100644
--- a/services/_shared/personas/code_reviewer/dispatch.py
+++ b/services/_shared/personas/code_reviewer/dispatch.py
@@ -48,7 +48,7 @@
     review_reasoner_diff,
 )
 from voice_pack import VoiceSelection, entitled_voice
-from review_types import EFFORTS
+from review_types import EFFORTS, SEVERITIES
 from personas.code_reviewer.dedup import (
     dedup_findings, finding_key, parse_rule, prior_keys_from_comments,
     rule_marker,
@@ -768,12 +768,14 @@ def _consolidated_agent_prompt(evaluation: CodeReviewEvaluation) -> str:
     return block
 
 
-# Derived from the shared vocabulary so a new effort level can never
-# silently drop its chip (the Severity-partition-assert drift class).
-_EFFORT_LABELS = {e: e.replace("-", " ") for e in EFFORTS}
-
 # Closed caveman-inspired chrome chips (FLINT-density scanability).
 # Identifiers and check *names* stay plain ASCII; Markings surface uses these.
+#
+# These two are hand-written, so they can drift from the shared vocabulary.
+# `_chip_vocabulary_drift()` is what stops that, and it is asserted in tests
+# against SEVERITIES/EFFORTS - a level added to review_types but not given a
+# chip here would otherwise fall through `.get()` to the bare identifier, and
+# nothing would say so.
 _SEVERITY_CHIP: dict[str, str] = {
     "critical": "💀 critical",
     "high": "🔥 high",
@@ -786,6 +788,21 @@ def _consolidated_agent_prompt(evaluation: CodeReviewEvaluation) -> str:
 }
 
 
+def _chip_vocabulary_drift() -> dict[str, frozenset[str]]:
+    """Vocabulary levels with no chip. Empty dict means no drift.
+
+    Returned rather than asserted at import so a drifted chip table degrades
+    to the bare identifier in production (ugly) instead of refusing to import
+    dispatch (fatal). The test is where it must fail loudly.
+    """
+    return {
+        k: v for k, v in {
+            "severity": SEVERITIES - set(_SEVERITY_CHIP),
+            "effort": EFFORTS - set(_EFFORT_CHIP),
+        }.items() if v
+    }
+
+
 def _severity_chip(severity: str) -> str:
     return _SEVERITY_CHIP.get(severity, severity)
 
diff --git a/services/webhook/tests/test_code_reviewer_dispatch.py b/services/webhook/tests/test_code_reviewer_dispatch.py
index 795c14a4..7b94b367 100644
--- a/services/webhook/tests/test_code_reviewer_dispatch.py
+++ b/services/webhook/tests/test_code_reviewer_dispatch.py
@@ -1950,11 +1950,28 @@ def test_agent_prompt_blocks_are_fence_breakout_safe():
     assert "````" in summary
 
 
-def test_effort_labels_derive_from_shared_vocabulary():
-    """A new effort level can never silently drop its chip - labels derive
-    from review_types.EFFORTS."""
-    from review_types import EFFORTS
-    assert frozenset(cr_dispatch._EFFORT_LABELS) == EFFORTS
+def test_chips_cover_the_whole_shared_vocabulary():
+    """A level added to review_types must not silently lose its chip.
+
+    This guard used to assert against `_EFFORT_LABELS`, a derived dict that
+    NOTHING used - so it could never drift and the test could never fail. The
+    dicts actually rendered, `_SEVERITY_CHIP` and `_EFFORT_CHIP`, are
+    hand-written and had no guard at all: a new severity would have fallen
+    through `.get()` to the bare identifier with nothing to say so. The guard
+    was watching a decoy.
+    """
+    assert cr_dispatch._chip_vocabulary_drift() == {}
+
+    # And it can actually fail - a guard nobody has seen go red is a guess.
+    from review_types import SEVERITIES
+    original = dict(cr_dispatch._SEVERITY_CHIP)
+    try:
+        cr_dispatch._SEVERITY_CHIP.pop(next(iter(SEVERITIES)))
+        assert cr_dispatch._chip_vocabulary_drift() != {}
+    finally:
+        cr_dispatch._SEVERITY_CHIP.clear()
+        cr_dispatch._SEVERITY_CHIP.update(original)
+    assert cr_dispatch._chip_vocabulary_drift() == {}