diff --git a/implementations/main.cpp b/implementations/main.cpp
index 1346d43..e3b735d 100644
--- a/implementations/main.cpp
+++ b/implementations/main.cpp
@@ -1230,7 +1230,12 @@ public:
     // store: an enrolment, or a template removed underneath us.
     int SetActiveGroup(std::uint32_t gid, bool force = false) {
         namespace ta = fingerprintd::ta;
-        if (!force && gid == gid_ && templatesLoaded_ >= 0) {
+        // > 0, never >= 0: caching a ZERO turns a failed load into a
+        // permanent one. It did exactly that on 2026-09-05 -- the trustlet
+        // unlinked a restored template, reported 0, and every later claim
+        // answered "skipped (already active, 0 template(s))" without ever
+        // trying again.
+        if (!force && gid == gid_ && templatesLoaded_ > 0) {
             std::println("SET_ACTIVE_GROUP gid={} skipped (already active, {} template(s))",
                          gid, templatesLoaded_);
             return templatesLoaded_;
diff --git a/packaging/fplearn.sh b/packaging/fplearn.sh
index a7e2adc..376b68f 100755
--- a/packaging/fplearn.sh
+++ b/packaging/fplearn.sh
@@ -2,8 +2,10 @@
 # fplearn.sh -- the template-learning measurement, as a protocol rather than a
 # pile of remembered commands.
 #
-#   fplearn.sh snapshot [name]     save the current template aside
-#   fplearn.sh restore  <name>     put a saved template back and reload it
+#   fplearn.sh snapshot [name]     save the current template aside (forensics
+#                                  only -- see restore)
+#   fplearn.sh restore  <name>     REFUSES: a restore DESTROYS the template on
+#                                  this device, see the note in the command
 #   fplearn.sh wipe                 remove every stored template (backup first)
 #   fplearn.sh enrol  [finger]     enrol at the config's sample count
 #   fplearn.sh base   [n] [w]      trial with learning OFF  (the baseline)
@@ -113,6 +115,34 @@ snapshot)
     echo "restore with: fplearn.sh restore $NAME" ;;
 
 restore)
+    # DISABLED, and it must stay disabled: restoring a template DESTROYS it.
+    #
+    # Measured 2026-09-05. A byte-exact restore of an earlier container made the
+    # trustlet read its header, UNLINK it, look for a .bak, and bump the RPMB
+    # counter -- for both containers. The enrolled finger was gone, and no copy
+    # can bring it back, because QTEE binds every SFS object to a hardware
+    # anti-rollback counter that has since advanced. An older container is not
+    # stale to QTEE, it is TAMPERED, and the response to tampering is deletion.
+    #
+    # So the paired within-session comparison this was built for is IMPOSSIBLE
+    # on this hardware. A template can be measured, and it can be replaced by
+    # enrolling again, but it can never be put back.
+    cat <<'WARN'
+REFUSING: restoring a template DESTROYS it on this device.
+
+QTEE binds every stored object to a hardware anti-rollback counter. An earlier
+container is not stale to it, it is tampered, and it responds by UNLINKING the
+file. Measured 2026-09-05: a byte-exact restore of the previous template made
+the trustlet unlink BOTH containers and advance the RPMB counter. The enrolled
+finger was destroyed and no copy could bring it back.
+
+Snapshots are still worth taking -- they are forensic evidence of what a
+template looked like -- but they cannot be put back. To change the template,
+enrol again.
+WARN
+    exit 1 ;;
+
+restore-force-DESTRUCTIVE)
     NAME=${2:-}
     D=$SNAPDIR/$NAME
     [ -n "$NAME" ] && [ -d "$D" ] || { echo "usage: fplearn.sh restore <name>"; echo "available:"; ls "$SNAPDIR" 2>/dev/null | sed 's/^/  /'; exit 1; }