diff --git a/sunnypilot/selfdrive/controls/lib/accel_personality/constants.py b/sunnypilot/selfdrive/controls/lib/accel_personality/constants.py
index 3764da223f..f81fb6a6fa 100644
--- a/sunnypilot/selfdrive/controls/lib/accel_personality/constants.py
+++ b/sunnypilot/selfdrive/controls/lib/accel_personality/constants.py
@@ -30,11 +30,15 @@ A_CRUISE_MAX_V = {
 }
 RISE_RATE = {ECO: 0.05, NORMAL: STOCK_RISE_RATE, SPORT: 0.06}   # ECO rise = stock so launch ramps promptly
 
-# Early soft braking: predicted brake need (m/s^2) -> early decel target (m/s^2).
+# Early soft braking: predicted brake need (m/s^2) -> early decel target (m/s^2). Front-loads a gentle
+# decel as soon as the 3s plan lookahead predicts a brake, so decel is spread out instead of arriving as
+# one late firm onset. The old ECO row was near-flat (-0.07 at brake_need~1.0 vs an eventual ~-0.88 plan
+# brake) so it barely front-loaded -> late, jerky onsets on route 00000456. Deepened toward (but kept
+# gentler than) NORMAL. Hard brakes (brake_need>=HARD_BRAKE_NEED or raw<=HARD_BRAKE_TARGET_ACCEL) still
+# bypass to stock, and min(.,raw) keeps it never weaker than the plan.
 SMOOTH_DECEL_BP = [0.0, 0.4, 0.8, 1.2, 1.6, 2.0, 2.4]
 SMOOTH_DECEL_V = {
-  ECO:    [0.00, -0.02, -0.05, -0.10, -0.25, -0.40, -0.60],
-  #ECO:    [0.00, -0.08, -0.20, -0.30, -0.40, -0.70, -0.80],
+  ECO:    [0.00, -0.08, -0.20, -0.35, -0.55, -0.78, -1.00],
   NORMAL: [0.00, -0.13, -0.30, -0.55, -0.84, -1.12, -1.40],
   SPORT:  [0.00, -0.17, -0.40, -0.72, -1.05, -1.35, -1.65],
 }
diff --git a/sunnypilot/selfdrive/controls/lib/radar_distance/radar_distance.py b/sunnypilot/selfdrive/controls/lib/radar_distance/radar_distance.py
index af3fde8b56..1ed1ea8656 100644
--- a/sunnypilot/selfdrive/controls/lib/radar_distance/radar_distance.py
+++ b/sunnypilot/selfdrive/controls/lib/radar_distance/radar_distance.py
@@ -16,7 +16,6 @@ from openpilot.common.realtime import DT_MDL
 HOLD_MAX_FRAMES = 10     # ~0.5s cap, measured since the last SUSTAINED lead (not reset by 1-frame flicker)
 SUSTAIN_FRAMES = 2       # consecutive valid frames to (re)arm and reset the wall-clock
 DROPOUT_DREL = 1.0
-MIN_PROB = 0.2
 FCW_PROB_CAP = 0.9       # held lead can't reach the FCW gate (>0.9) -> no false FCW
 MIN_HELD_DREL = 0.5
 
@@ -56,7 +55,11 @@ class _LeadHold:
     self.__init__()
 
   def step(self, raw):
-    if raw.status and raw.dRel > DROPOUT_DREL and raw.modelProb > MIN_PROB:
+    # Validity mirrors the MPC, which keys off status alone (long_mpc process_lead). modelProb is NOT a
+    # gate: radard's low_speed_override emits a real closest-track lead with modelProb=0.0, so gating on
+    # prob wrongly rejected real close stop-and-go leads and substituted a stale farther held lead ->
+    # under-brake -> stopping too close. FCW stays bounded by FCW_PROB_CAP on the held output below.
+    if raw.status and raw.dRel > DROPOUT_DREL:
       self._last = (raw.dRel, raw.vRel, raw.vLead, raw.aLeadK, raw.aLeadTau, raw.modelProb)
       self._sustained += 1
       if self._sustained >= SUSTAIN_FRAMES:
diff --git a/sunnypilot/selfdrive/controls/lib/radar_distance/tests/test_radar_distance.py b/sunnypilot/selfdrive/controls/lib/radar_distance/tests/test_radar_distance.py
index 8029c5c2a6..87e5777611 100644
--- a/sunnypilot/selfdrive/controls/lib/radar_distance/tests/test_radar_distance.py
+++ b/sunnypilot/selfdrive/controls/lib/radar_distance/tests/test_radar_distance.py
@@ -64,6 +64,25 @@ def test_holds_after_sustained_dropout():
   assert held.dRel == pytest.approx(30.0 - 4.0 * 0.05, abs=1e-6)
 
 
+def test_low_speed_override_lead_passthrough():
+  # radard low_speed_override emits a real closest-track lead with modelProb=0.0. It must be honored as a
+  # real lead (passthrough), NOT rejected and replaced by a stale farther held lead (would under-brake at
+  # stop-and-go and stop too close).
+  c = ctrl()
+  one = lead(status=True, dRel=2.5, vRel=0.0, vLead=0.0, modelProb=0.0)
+  out = c.smooth_radarstate(rs(one))
+  assert out.leadOne is one                         # passed straight through, not substituted
+
+
+def test_low_speed_override_lead_arms_hold():
+  # a sustained prob=0 real lead should arm the hold like any real lead
+  c = ctrl()
+  for _ in range(3):
+    c.smooth_radarstate(rs(lead(status=True, dRel=3.0, vRel=-0.5, vLead=1.0, modelProb=0.0)))
+  held = c.smooth_radarstate(rs(lead(status=False, dRel=0.0, modelProb=0.0))).leadOne
+  assert held.status is True                         # armed off the prob=0 lead, holds through dropout
+
+
 def test_obstacle_monotone_during_hold():
   c = ctrl()
   for _ in range(3):