diff --git a/goodwe_controller/CHANGELOG.md b/goodwe_controller/CHANGELOG.md
index 331eb36..d0a9970 100644
--- a/goodwe_controller/CHANGELOG.md
+++ b/goodwe_controller/CHANGELOG.md
@@ -4,7 +4,7 @@
 
 ### Not in a release yet
 
-- The night grid charge no longer raises "Battery stopped limiting the peak" when the 15-minute period stays on the aim ([[SAFETY-126]]).
+- The night grid charge no longer raises "Battery stopped limiting the peak" when the 15-minute period stays on the aim. A steadily rising load now alarms a little later, and with a peak margin of 0 a period can end up to 10 W over the limit without this alarm ([[SAFETY-126]]).
 
 **A patch bump: one fix for charging from the grid under the peak limit, and no option changes.**
 No option was added, removed or renamed, and no default changed (86 options).
diff --git a/goodwe_controller/DECISIONS/SAFETY-126-landing-error-is-not-an-overrun.md b/goodwe_controller/DECISIONS/SAFETY-126-landing-error-is-not-an-overrun.md
index a2099c4..feda5a4 100644
--- a/goodwe_controller/DECISIONS/SAFETY-126-landing-error-is-not-an-overrun.md
+++ b/goodwe_controller/DECISIONS/SAFETY-126-landing-error-is-not-an-overrun.md
@@ -25,16 +25,21 @@ no "Battery stopped limiting the peak" event hint. The claim, the floor and the
 commands are unchanged.
 
 **Still alarms.** House load the batteries cannot cover under the reserve:
-the period would end over the aim with the batteries at 0 W, so the alarm
-fires on the same cycle as before (both under the reserve; a 6 500 W house
-during the night charge). SAFETY-87's case (others able, graded on the limit
-for `ALARM_HOLD_S`) and SAFETY-81's 'Load too high' notification are unchanged.
-A period heading at most 10 W over the aim no longer alarms with nothing able
-to give. That is SAFETY-81's landing tolerance. `misses_aim()` takes the house
-as meter plus batteries, so an export nets off. `decide()`'s projection counts
-offtake only. So a re-alarm while the loop's draw on held batteries covers the
-house with some export, and the period lands under the aim, is gone too
-(test_safety126.py part 2, modes off: 2 notifications, was 4).
+the period would end over the aim with the batteries at 0 W. A step load
+alarms on the same cycle as before. SAFETY-87's case and SAFETY-81's 'Load too
+high' notification are unchanged.
+
+**What the 10 W tolerance costs** (review, owner accepted 2026-10-31):
+- A steadily rising load alarms a little later, once the projected miss passes
+  10 W (3 000 -> 4 000 W over 45 min: 56 s later). With the default 50 W
+  margin the alarm still comes before the limit.
+- With `peak_margin_w` 0, a period can end up to 10 W over the limit without
+  this alarm, the tolerance 'Load too high' already accepts (SAFETY-81).
+
+`misses_aim()` takes the house as meter plus batteries, so an export nets
+off (`decide()`'s projection counts offtake only). A re-alarm while the loop's
+draw on held batteries covers the house is gone too (test_safety126.py part 2,
+modes off: 2 notifications, was 4).
 
 **Not changed.** The one-cycle floor itself: at a period's last instant the
 guard still asks an able battery for its rating, or stops a charge, for a
diff --git a/goodwe_controller/DOCS.md b/goodwe_controller/DOCS.md
index f055592..2f08c2f 100644
--- a/goodwe_controller/DOCS.md
+++ b/goodwe_controller/DOCS.md
@@ -3458,7 +3458,7 @@ Charging from the grid stays blocked meanwhile. With several inverters the
 notification names the battery at its limit (SAFETY-32), each graded against
 its own cut-out level. That battery is held (weight 0 while the site gives)
 and the guard's need goes to the others, up to their rating; the alarm fires
-only if they miss the limit at full power for `ALARM_HOLD_S` (SAFETY-87). With no battery able to give it fires at once, but only if the period would end more than 10 W over the aim with the batteries at 0 W. A grid charge the add-on runs counts as one it can stop, and the last second's landing error on the aim is not an alarm (SAFETY-126). ⚠️ **The reserve binds the peak guard only
+only if they miss the limit at full power for `ALARM_HOLD_S` (SAFETY-87). With no battery able to give it fires at once, but only if the period would end more than 10 W over the aim with the batteries at 0 W. A grid charge the add-on runs counts as one it can stop, and the last second's landing error on the aim is not an alarm. A steadily rising load therefore alarms a little later. With `peak_margin_w` 0 a period can end up to 10 W over the limit without this alarm, as for “Load too high” (SAFETY-126). ⚠️ **The reserve binds the peak guard only
 (SAFETY-25).** Self-consumption has no such floor. See *What `peak_soc_floor`
 does not protect*.
 
diff --git a/goodwe_controller/test_safety126.py b/goodwe_controller/test_safety126.py
index a76f152..5d3275c 100644
--- a/goodwe_controller/test_safety126.py
+++ b/goodwe_controller/test_safety126.py
@@ -18,8 +18,9 @@ quarter's edge as on the rig.
    red peak row, no "Why" peak, every quarter on the aim; the guard did want
    power at the quarter's end (not vacuous); with the fix reverted the same
    run alarms. The commands are the same with and without the fix.
-2. A real overrun under the reserve still alarms, at the same cycle as
-   before, and names both batteries.
+2. A real overrun under the reserve still alarms, a step load at the same
+   cycle as before, and names both batteries. After it, the latch is reset:
+   the night's later landings are quiet again.
 3. `misses_aim()` on its own.
 4. The page and the "Why" on a decision within the aim.
 
@@ -345,6 +346,48 @@ for tag, onset, kw in (
           any(s[4] for s in head["seen"]))
 
 
+def burst(t):
+    """600 W, a 6 500 W load from 20 to 25 minutes, then 600 W again while
+    the night charge goes on."""
+    return 6500.0 if 1200 <= t < 1500 else 600.0
+
+
+# F3 (review): `_peak_landing()` resets its latch when the guard lets go, so
+# a real overrun does not keep the later landing errors alarming.
+_bh = drive(opts(), house=burst, off=0.99, minutes=61)
+_bb = without_fix(lambda: drive(opts(), house=burst, off=0.99,
+                                     minutes=61))
+show("head real overrun, then the night again", _bh)
+show("fix reverted", _bb)
+# The burst banked a new month peak, which the later quarters stay under.
+_new_peak = _bh["c"].peak_guard.ceiling_w
+
+
+def _late(r):
+    """From the next quarter on: the overrun's own quarter ends over."""
+    return [x for x in r["seen"] if x[0] >= 1800]
+
+
+print(f"    notifications at {[t for t, _n in _bh['notes']]} s; with the fix "
+      f"reverted {[t for t, _n in _bb['notes']]} s; limit after the burst "
+      f"{_new_peak:.0f} W")
+check("⭐ after a real overrun ends, the latch is reset: the alarm came "
+      "during the overrun, and the later quarters' landings are quiet "
+      "(no notification, no 'Why' peak, no red row) and under the new peak",
+      any(1200 <= t < 1500 for t, _n in _bh["notes"])
+      and not [t for t, _n in _bh["notes"] if t >= 1800]
+      and not any(x[3] or x[4] or x[5] for x in _late(_bh))
+      and present(nth([v for k, v in _bh["quarters"].items()
+                       if k >= "22:30"], 0))
+      and all(v <= _new_peak for k, v in _bh["quarters"].items()
+              if k >= "22:30"))
+check("...not vacuous: the guard wanted power at those landings, and with "
+      "the fix reverted they show the 'Why' peak or an alarm",
+      any(x[2] for x in _late(_bh))
+      and (any(x[3] or x[4] or x[5] for x in _late(_bb))
+           or bool([t for t, _n in _bb["notes"] if t >= 1800])))
+
+
 # ============================================================================ #
 print("\n3. misses_aim() on its own")
 START = datetime(2026, 10, 30, 21, 15, tzinfo=timezone.utc)