Add gibberish-scores-well test: proves F3.3 is anti-spam not quality metric (16/16)

test_swarmmetrics.py

@@ -188,6 +188,36 @@ def test_f32_idf_echo_combined():
     assert edge.reciprocity_f32 > 0, "F3.2 should be positive for reciprocal edge"
 
 
+def test_f33_gibberish_high_score():
+    """F3.3 CANNOT detect quality — only filter spam.
+    High-reciprocity, high-diversity gibberish scores well.
+    This is not a bug — it's the boundary of what statistics can measure."""
+    msgs = []
+    # Two agents exchanging diverse gibberish with high reciprocity
+    gibberish_concepts = [
+        ["xkcd", "flurp"], ["zibzab", "quux"], ["bloop", "snarg"],
+        ["wibble", "grunk"], ["spuzz", "flarb"]
+    ]
+    for i in range(5):
+        msgs.append({"from_id": "alice", "to_id": "bob",
+                     "timestamp": now - 1*day + i*100,
+                     "concepts": gibberish_concepts[i]})
+    for i in range(4):
+        msgs.append({"from_id": "bob", "to_id": "alice",
+                     "timestamp": now - 0.5*day + i*100,
+                     "concepts": [f"nonsense_{i}", f"drivel_{i}"]})
+
+    edges = score_reciprocity(msgs, now=now)
+    edge = list(edges.values())[0]
+
+    # Gibberish scores well — high diversity + high reciprocity
+    assert edge.reciprocity_f33 > 0.5, \
+        f"Gibberish should score high on F3.3: {edge.reciprocity_f33:.4f}"
+    # This is the proof: F3.3 measures not-spam, not quality.
+    # 18 unique concepts, reciprocity ~0.8 → high score.
+    # The formula chain is an anti-spam filter, not a quality metric.
+
+
 # Run all tests
 tests = [v for k, v in sorted(globals().items()) if k.startswith("test_")]
 passed = 0