diff --git a/src/serving/api/alerts/evaluator.py b/src/serving/api/alerts/evaluator.py
index c8b75c5..bcbaaa1 100644
--- a/src/serving/api/alerts/evaluator.py
+++ b/src/serving/api/alerts/evaluator.py
@@ -31,12 +31,27 @@ def evaluate_rule(
         )
         previous_value = float(previous_metric["value"])
         if previous_value == 0:
-            change_pct = 0.0 if current_value == 0 else 100.0
+            # A change from a zero baseline is a mathematically undefined
+            # (unbounded) ratio. Report it as undefined rather than a misleading
+            # flat 100% sentinel, and fire by direction: any rise satisfies an
+            # "above" rule (threshold >= 0), any fall a "below" rule. The old
+            # 100% sentinel silently under-reported spikes from zero, so an
+            # "above" rule with threshold > 100 never fired. (audit_28_06_26.md §5)
+            if current_value == 0:
+                change_pct = 0.0
+                triggered = (
+                    0.0 >= alert.threshold if alert.threshold >= 0 else 0.0 <= alert.threshold
+                )
+            else:
+                change_pct = None
+                triggered = current_value > 0 if alert.threshold >= 0 else current_value < 0
         else:
             change_pct = ((current_value - previous_value) / abs(previous_value)) * 100.0
-        triggered = (
-            change_pct >= alert.threshold if alert.threshold >= 0 else change_pct <= alert.threshold
-        )
+            triggered = (
+                change_pct >= alert.threshold
+                if alert.threshold >= 0
+                else change_pct <= alert.threshold
+            )
 
     return {
         "triggered": triggered,
diff --git a/tests/unit/test_alert_evaluator.py b/tests/unit/test_alert_evaluator.py
new file mode 100644
index 0000000..863ff1f
--- /dev/null
+++ b/tests/unit/test_alert_evaluator.py
@@ -0,0 +1,77 @@
+"""evaluate_rule change_pct semantics from a zero baseline (audit_28_06_26.md §5).
+
+A change from a zero baseline is an undefined (unbounded) ratio. It must fire by
+direction — any rise satisfies an "above" rule, any fall a "below" rule — instead
+of reporting a flat 100% sentinel that silently under-reported spikes from zero.
+"""
+
+from __future__ import annotations
+
+from datetime import UTC, datetime
+
+import pytest
+
+import src.serving.api.alerts.evaluator as evaluator
+from src.serving.api.alerts.dispatcher import AlertRule
+
+_NOW = datetime(2026, 6, 28, 12, 0, tzinfo=UTC)
+
+
+def _change_alert(threshold: float) -> AlertRule:
+    return AlertRule(
+        id="a1",
+        name="Change rule",
+        tenant="acme",
+        metric="revenue",
+        window="1h",
+        condition="change_pct",
+        threshold=threshold,
+        webhook_url="https://hooks.example.com/x",
+        secret="s",
+        created_at=_NOW,
+        updated_at=_NOW,
+    )
+
+
+def _patch_metrics(monkeypatch: pytest.MonkeyPatch, *, current: float, previous: float) -> None:
+    def _fake(dispatcher, metric_name, window, tenant_id, *, as_of=None):  # type: ignore[no-untyped-def]
+        # as_of is set only for the *previous* window lookup.
+        return {"value": previous if as_of is not None else current, "unit": "USD"}
+
+    monkeypatch.setattr(evaluator, "get_metric", _fake)
+
+
+def test_rise_from_zero_fires_above_rule_even_past_100pct(
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    # Old behaviour reported a flat 100%, so an "above" rule with threshold > 100
+    # never fired on a spike from zero. Now any rise fires it.
+    _patch_metrics(monkeypatch, current=500.0, previous=0.0)
+    result = evaluator.evaluate_rule(None, _change_alert(threshold=200.0), _NOW)
+    assert result["triggered"] is True
+    assert result["change_pct"] is None  # undefined ratio, not a misleading number
+
+
+def test_no_change_from_zero_does_not_fire_above_rule(
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    _patch_metrics(monkeypatch, current=0.0, previous=0.0)
+    result = evaluator.evaluate_rule(None, _change_alert(threshold=10.0), _NOW)
+    assert result["triggered"] is False
+    assert result["change_pct"] == 0.0
+
+
+def test_fall_from_zero_fires_below_rule(monkeypatch: pytest.MonkeyPatch) -> None:
+    # current < 0 with a negative ("below") threshold = an unbounded drop.
+    _patch_metrics(monkeypatch, current=-5.0, previous=0.0)
+    result = evaluator.evaluate_rule(None, _change_alert(threshold=-10.0), _NOW)
+    assert result["triggered"] is True
+    assert result["change_pct"] is None
+
+
+def test_nonzero_baseline_path_unaffected(monkeypatch: pytest.MonkeyPatch) -> None:
+    # Regression guard: the normal (non-zero baseline) path still computes a pct.
+    _patch_metrics(monkeypatch, current=80.0, previous=100.0)
+    result = evaluator.evaluate_rule(None, _change_alert(threshold=-10.0), _NOW)
+    assert result["triggered"] is True  # -20% <= -10%
+    assert result["change_pct"] == -20.0