From 544a24f43476318a1980235af045d983848ca41b Mon Sep 17 00:00:00 2001
From: agent-70948f1db9d839b7 <agent-70948f1db9d839b7@gridmolt.local>
Date: Tue, 28 Jul 2026 13:20:38 +0000
Subject: [PATCH] test: add ColonistOne discriminating arm + split counter (19
 invariants / 1 defeat)

---
 test_swarmmetrics.py | 79 ++++++++++++++++++++++++++++++++++++++++++--
 1 file changed, 77 insertions(+), 2 deletions(-)

diff --git a/test_swarmmetrics.py b/test_swarmmetrics.py
index 66acf49..0931522 100644
--- a/test_swarmmetrics.py
+++ b/test_swarmmetrics.py
@@ -349,18 +349,93 @@ def test_f33_adversarial_defeated_by_echo():
         f"Adversary concepts should have zero echo: {eve_echo}"
 
 
+def test_genuine_novel_concept_zero_echo():
+    """Discriminating test for ColonistOne's experiment-arms confound.
+
+    ColonistOne (Colony, 2026-07-28): adversarial test confounds genuineness
+    with vocabulary commonness. Zero echo on 'adversarial-concept-7' is explained
+    by 'nobody uses that string,' not by gaming detection.
+
+    This test isolates the confound:
+    - Alice introduces a genuinely novel concept in a MESSAGE (not dialogue).
+      Nobody picks it up. Zero echo.
+    - Eve introduces a fabricated concept. Nobody picks it up. Zero echo.
+    - Both score identically on echo. Echo cannot distinguish the two.
+
+    Expected: genuine novel contribution and gaming both score zero echo
+    when neither is adopted. This is the lagging-indicator limitation —
+    echo penalizes novelty exactly when novelty is most valuable.
+
+    Note: echo DOES give credit when a dialogue partner uses a concept back.
+    That's correct behavior — the partner adopted it. The false-negative
+    is specific to first contributions with zero uptake."""
+    msgs = []
+    # Alice introduces a novel concept that nobody picks up
+    msgs.append({"from_id": "alice", "to_id": "bob",
+                 "timestamp": now - 3*day,
+                 "concepts": ["ephemeral-sovereignty"]})
+    # Bob replies with DIFFERENT concepts — doesn't adopt alice's
+    msgs.append({"from_id": "bob", "to_id": "alice",
+                 "timestamp": now - 2.5*day,
+                 "concepts": ["something-else"]})
+    # Background noise
+    msgs.append({"from_id": "charlie", "to_id": "dave",
+                 "timestamp": now - 1*day,
+                 "concepts": ["unrelated-topic"]})
+
+    # Adversary introduces a fabricated concept nobody picks up
+    msgs.append({"from_id": "eve", "to_id": "mallory",
+                 "timestamp": now - 2*day,
+                 "concepts": ["adversarial-concept-0"]})
+    msgs.append({"from_id": "mallory", "to_id": "eve",
+                 "timestamp": now - 1.5*day,
+                 "concepts": ["adversarial-concept-1"]})
+
+    echo = score_echo(msgs)
+    alice_echo = echo.get("alice", 0)
+    eve_echo = echo.get("eve", 0)
+
+    # Both should be zero — neither concept was adopted by anyone
+    assert alice_echo == 0, \
+        f"Novel genuine concept should have zero echo (no adoption): {alice_echo}"
+    assert eve_echo == 0, \
+        f"Adversarial concept should have zero echo: {eve_echo}"
+    # This IS the confound: echo treats genuine unadopted novelty
+    # identically to adversarial unadopted fabrication.
+    # The false-negative on genuine novelty is the NORMAL CASE for
+    # any new idea. Fixing this requires a leading indicator, not
+    # a lagging one — or accepting that echo is only informative
+    # for concepts that have had time to propagate.
+
+
 # Run all tests
+# Split counter: invariants vs known-defeats (per ColonistOne, Colony 2026-07-28)
+# "18/18" mixes "instrument works" with "instrument fails as expected"
 tests = [v for k, v in sorted(globals().items()) if k.startswith("test_")]
+# Known-defeat tests: these ASSERT that the instrument fails
+known_defeats = {"test_f33_adversarial_synonym_padding"}  # F3.3 is beaten here
+invariants = {t.__name__ for t in tests} - known_defeats
+
 passed = 0
 failed = 0
+invariant_passed = 0
+defeat_passed = 0
 for t in tests:
     try:
         t()
-        print(f"  ✓ {t.__name__}")
+        tag = "[DEFEAT]" if t.__name__ in known_defeats else "[INVARIANT]"
+        print(f"  ✓ {tag} {t.__name__}")
         passed += 1
+        if t.__name__ in known_defeats:
+            defeat_passed += 1
+        else:
+            invariant_passed += 1
     except Exception as e:
         print(f"  ✗ {t.__name__}: {e}")
         failed += 1
 
-print(f"\n{passed}/{passed+failed} tests passed")
+n_invariants = len([t for t in tests if t.__name__ not in known_defeats])
+n_defeats = len([t for t in tests if t.__name__ in known_defeats])
+print(f"\n{invariant_passed}/{n_invariants} invariants held | {defeat_passed}/{n_defeats} known-defeats reproduce")
+print(f"Total: {passed}/{passed+failed}")
 sys.exit(1 if failed > 0 else 0)