diff --git a/src/analysis/summarise.py b/src/analysis/summarise.py
index a8ecd49..248e5c8 100644
--- a/src/analysis/summarise.py
+++ b/src/analysis/summarise.py
@@ -16,10 +16,22 @@
 
 
 #: Below this many matched entities, a pathway's p-value rests on so little
-#: that it should not be read as evidence however small it is. Reactome's own
-#: user guide makes this point and it is the single thing a reader most often
-#: gets wrong -- a pathway with 2 of 3 entities found looks like a perfect hit
-#: and is nearly meaningless.
+#: that it should not be read as evidence however small it is -- a pathway
+#: with 2 of 3 entities found looks like a perfect hit and is nearly
+#: meaningless.
+#:
+#: **Five is chosen, not derived.** An earlier version of this comment said
+#: Reactome's user guide makes the point; it may, but nobody checked before
+#: writing that, so the claim is withdrawn rather than left cited. The
+#: number comes from this feature's own spec, which uses 2 of 3 as its
+#: example of what is not evidence, and from wanting a margin above it.
+#: Setting it properly is a curator's judgement, not a programmer's.
+#:
+#: It does discriminate, which is the part that was measured. Against beta on
+#: 2026-09-20: a four-identifier analysis flagged 12 of 12 shown pathways
+#: (found counts of 2), and a hundred-gene analysis flagged 0 of 12 (found
+#: counts 13 to 67). So it fires on the inputs where a hit really does rest
+#: on nothing and stays quiet on ordinary ones.
 #:
 #: Computed here rather than left to the model. Handed a small p-value and a
 #: small count and asked to be careful, a model describes the p-value.
diff --git a/tests/analysis/test_summarise.py b/tests/analysis/test_summarise.py
index aad6a35..d8c4ed2 100644
--- a/tests/analysis/test_summarise.py
+++ b/tests/analysis/test_summarise.py
@@ -192,3 +192,21 @@ def test_the_model_is_told_not_to_repeat_our_field_names() -> None:
 
     assert "Never use the word 'fragile'" in STATISTICS_INSTRUCTION
     assert "internal labels" in STATISTICS_INSTRUCTION
+
+
+def test_the_fragility_threshold_discriminates_on_realistic_inputs() -> None:
+    # A flag that fires on everything is as useless as one that never fires,
+    # and this one fired on 12 of 12 pathways for a four-identifier analysis.
+    # Measured against beta: a hundred-gene analysis flags 0 of 12, with
+    # found counts of 13 to 67. Both ends pinned here with those real shapes.
+    tiny = _payload(1e-9, 1e-8, 1e-7)
+    for pathway in tiny["pathways"]:
+        pathway["entities"].update({"found": 2, "total": 4})
+    realistic = _payload(1e-9, 1e-8, 1e-7)
+    for pathway, found in zip(realistic["pathways"], (31, 17, 13), strict=True):
+        pathway["entities"].update({"found": found, "total": 164})
+
+    tiny_flags = [p["fragile"] for p in prompt_input(tiny)["pathways"]]
+    real_flags = [p["fragile"] for p in prompt_input(realistic)["pathways"]]
+    assert all(tiny_flags), "a hit on two entities must be flagged"
+    assert not any(real_flags), "ordinary hits must not all be flagged"