diff --git a/packaging/fplearn.sh b/packaging/fplearn.sh
index 7c3ae9a..ebc4032 100755
--- a/packaging/fplearn.sh
+++ b/packaging/fplearn.sh
@@ -2,12 +2,26 @@
 # 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 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)
 #   fplearn.sh trend  [n] [w]      three trials with learning ON
 #   fplearn.sh sizes                just print the template container sizes
 #
+# SNAPSHOT BEFORE EVERY MEASUREMENT BLOCK. Learning rewrites the template in
+# place, so without a snapshot an improvement cannot be attributed: a rate that
+# climbs over an evening is equally explained by the template getting better, by
+# the SENSOR being wiped clean, or by the USER learning where the sensor likes
+# to be pressed (Jorijn, 2026-09-04, having spotted both while the numbers were
+# still going up). All three move the same way on the same time axis.
+#
+# A snapshot makes the paired test possible: measure the learned template, then
+# restore the older one and measure again WITHIN THE SAME SESSION, so the sensor
+# is equally clean and the user equally practised for both halves. A rate that
+# falls back on restore is learning. A rate that stays up is not.
+#
 # WIPE FIRST when re-enrolling a finger that is already in the group. Measured
 # 2026-09-04: the duplicated-finger check refuses a finger the group already
 # holds -- 0 accepted of 7 presses, rc=0 on every one, while a never-enrolled
@@ -76,7 +90,46 @@ restart() {  # $1 = extra args
     printf ' TIMED OUT\n'; return 1
 }
 
+SNAPDIR=$HOME/fp6-backups/snapshots
+
 case "${1:-}" in
+snapshot)
+    NAME=${2:-$(date +%Y%m%d-%H%M%S)}
+    D=$SNAPDIR/$NAME
+    mkdir -p "$D"
+    n=0
+    for f in $(sudo ls "$GROUP"); do
+        sz=$(sudo stat -c %s "$GROUP/$f")
+        if [ "$sz" -gt 200000 ]; then sudo cp -p "$GROUP/$f" "$D/$f"; n=$((n+1)); fi
+    done
+    sudo cp -p "$MAP" "$D/" 2>/dev/null || true
+    sudo chown -R "$(id -u):$(id -g)" "$D"
+    echo "snapshot '$NAME': $n container(s)"
+    ls -la "$D" | awk 'NR>3{printf "  %9d  %s\n", $5, $9}'
+    echo "restore with: fplearn.sh restore $NAME" ;;
+
+restore)
+    NAME=${2:-}
+    D=$SNAPDIR/$NAME
+    [ -n "$NAME" ] && [ -d "$D" ] || { echo "usage: fplearn.sh restore <name>"; echo "available:"; ls "$SNAPDIR" 2>/dev/null | sed 's/^/  /'; exit 1; }
+    echo "=== restore snapshot '$NAME' ==="
+    echo "before:"; sizes
+    # The trustlet holds the template in memory once loaded, so a restore that
+    # does not restart the daemon changes the file and nothing else.
+    sudo systemctl stop "$UNIT" 2>/dev/null; sleep 2
+    for f in $(ls "$D"); do
+        case "$f" in fingers-*.map) sudo cp -p "$D/$f" "$MAP" ;;
+                     *) sudo cp -p "$D/$f" "$GROUP/$f" ;; esac
+    done
+    sudo sync
+    restart "--learn=0" || exit 1
+    echo "after:"; sizes
+    fprintd-list user 2>&1 | tail -1
+    echo
+    echo "Daemon is up with learning OFF, so the restored template stays put."
+    echo "Measure it now, in this session, against the block you just ran:"
+    echo "  fplearn.sh base" ;;
+
 sizes)
     echo "template containers now:"; sizes ;;
 
@@ -156,5 +209,5 @@ trend)
     echo "does not is noise, and the daemon's own 'learn:' lines say which." ;;
 
 *)
-    sed -n '2,38p' "$0"; exit 1 ;;
+    sed -n '2,52p' "$0"; exit 1 ;;
 esac