diff --git a/src/quant_strategy_plugins/ai_audit.py b/src/quant_strategy_plugins/ai_audit.py
index 588d73c..4e6d412 100644
--- a/src/quant_strategy_plugins/ai_audit.py
+++ b/src/quant_strategy_plugins/ai_audit.py
@@ -2,6 +2,7 @@
 
 import json
 import logging
+import math
 import os
 import re
 from collections.abc import Callable, Mapping, Sequence
@@ -444,6 +445,8 @@ def _as_confidence(value: Any) -> float | None:
         confidence = float(value)
     except (TypeError, ValueError):
         return None
+    if not math.isfinite(confidence):
+        return None
     if confidence < 0.0:
         return 0.0
     if confidence > 1.0:
@@ -699,7 +702,7 @@ def _run_ai_audit(
             attempts.append({**endpoint.report(), "status": status})
 
             # Advisory content is display-only; it must not mutate feedback state.
-            if not advisory:
+            if not advisory and audit_response.get("confidence") is not None:
                 _report_shadow_disagreement(
                     audit_kind=audit_kind,
                     ai_verdict=audit_response.get("verdict", ""),
diff --git a/tests/test_ai_audit.py b/tests/test_ai_audit.py
index 72e9576..7ce448e 100644
--- a/tests/test_ai_audit.py
+++ b/tests/test_ai_audit.py
@@ -1,5 +1,6 @@
 import sys
 import types
+import json
 
 import pytest
 
@@ -15,6 +16,31 @@
 )
 
 
+@pytest.mark.parametrize("confidence", ["nan", "NaN", "inf", "-inf", float("nan"), float("inf"), -float("inf")])
+@pytest.mark.parametrize("entry", [ai_audit.run_crisis_ai_audit, ai_audit.run_taco_ai_audit])
+def test_nonfinite_confidence_remains_unknown_and_does_not_write_feedback(monkeypatch, confidence, entry):
+    monkeypatch.setenv("CODEX_AUDIT_SERVICE_URL", "https://gateway.invalid")
+    monkeypatch.setattr(ai_audit, "build_ai_audit_endpoints", lambda **_: (
+        ai_audit.AiAuditEndpoint("primary", "", model="synthetic-model"),
+    ))
+    monkeypatch.setattr(ai_audit, "_complete_with_endpoint", lambda *_: (
+        json.dumps({"verdict": "agree", "confidence": confidence}), False,
+    ))
+    feedback = []
+    monkeypatch.setattr(ai_audit, "_report_shadow_disagreement", lambda **fields: feedback.append(fields))
+    payload = entry({"canonical_route": "no_action"}, enabled=True)
+    assert payload["confidence"] is None
+    assert feedback == []
+    assert payload["final_route_unchanged"] is True
+    assert payload["execution_controls"]["broker_order_allowed"] is False
+    json.dumps(payload, allow_nan=False)
+
+
+@pytest.mark.parametrize("value,expected", [(None, None), ("invalid", None), (0.0, 0.0), ("0.8", 0.8), (1.0, 1.0)])
+def test_confidence_finite_and_missing_values_keep_existing_semantics(value, expected):
+    assert ai_audit._as_confidence(value) == expected
+
+
 def _clear_ai_audit_env(monkeypatch) -> None:
     for key in (
         "QSP_STRATEGY_PLUGIN_AI_AUDIT_API_KEY",