diff --git a/static/cosmopolite.js b/static/cosmopolite.js
index c4ae9c3..4d94727 100644
--- a/static/cosmopolite.js
+++ b/static/cosmopolite.js
@@ -171,6 +171,10 @@ Cosmopolite.prototype.onLoad_ = function() {
   if (--this.numScriptsToLoad_ > 0) {
     return;
   }
+  // jQuery.noConflict() doesn't actually remove window.$ and window.jQuery,
+  // it just sets them back to undefined. This angers QUnit's globals
+  // detection. goog.appengine.Channel doesn't even provide a noConflict()
+  // equivalent.
   this.$ = jQuery.noConflict(true);
   if (this.shutdown_) {
     // Shutdown during startup
diff --git a/static/test.js b/static/test.js
index ba2bf16..68433db 100644
--- a/static/test.js
+++ b/static/test.js
@@ -25,6 +25,17 @@ server interactions are complex and tests should be structured to verify them,
 not to verify the behavior of a simulation.
 */
 
+/*
+These tests break if you turn on global pollution detection because of at
+least:
+
+* $, jQuery: jQuery's noConflict() doesn't actually delete $ or jQuery; it
+    sets them to undefined.
+* goog: goog.appengine.Channel doesn't provide any kind of noConflict()
+    equivalent.
+* closure_lm_*: The Channel code has a bug that puts this in globals.
+*/
+
 var randstring = function() {
   var ret = [];
   for (var i = 0; i < 16; i++) {