diff --git a/tools/fleet/autorun.py b/tools/fleet/autorun.py
index ae9689da1..fdbbc0ad2 100644
--- a/tools/fleet/autorun.py
+++ b/tools/fleet/autorun.py
@@ -27,9 +27,11 @@ def start(b, script, env=""):
 last_publish = 0; retried = set()
 while True:
     reg = fleet.load(); changed = False
-    for iid, b in list(reg.items()):
-        if b.get("state") in ("destroyed", "failed") or not b.get("ssh_ok") or b.get("phase") == "5": continue
-        p = probe(b)
+    live = [(iid, b) for iid, b in reg.items() if b.get("state") not in ("destroyed", "failed") and b.get("ssh_ok") and b.get("phase") != "5"]
+    from concurrent.futures import ThreadPoolExecutor
+    with ThreadPoolExecutor(max_workers=16) as ex: probes = dict(zip([i for i, _ in live], ex.map(lambda x: probe(x[1]), live)))
+    for iid, b in live:
+        p = probes.get(iid)
         if p is None: continue
         stage = b.get("stage", "setup")
         if p["last"]: fleet.patch(iid, last_line=p["last"])
diff --git a/tools/fleet/box-kill.sh b/tools/fleet/box-kill.sh
new file mode 100755
index 000000000..5c680acd5
--- /dev/null
+++ b/tools/fleet/box-kill.sh
@@ -0,0 +1,15 @@
+#!/usr/bin/env bash
+# Stops every stage process on a box (the matrix, Ember, the prover loop, their miners, workers, samplers and SP1
+# servers); never the node. Run as a FILE (bash in/box-kill.sh) so that no pattern below can match the shell that runs
+# it (an inline `pkill -f 'igneum-prove-host'` killed its own ssh shell on 6 October 2026, 12:38Z).
+for i in 1 2; do
+  pkill -9 -f '^bash in/box-matrix.sh' ; pkill -9 -f '^bash in/box-ember.sh'; pkill -9 -f '^bash in/box-prover.sh'; pkill -9 -f '^python3 -u /root/fleet/in/box-prover.py'
+  pkill -9 -x igneum-miner; pkill -9 -f '^/opt/igneum/pkg/bin/igneum-worker-cuda'; pkill -9 -x sp1-gpu-server; pkill -9 -f '^/opt/igneum-floor/(bin|target|prove)/.*igneum-prove-(host|export)'   # pkill -x cannot match a name over 15 characters (igneum-worker-cuda, igneum-prove-host)
+  pkill -9 -f '^nvidia-smi --query'
+  sleep 2
+done
+rm -f /tmp/sp1-cuda-*.sock
+nvidia-smi -i 0 -rgc >/dev/null 2>&1
+d="$(nvidia-smi --query-gpu=power.default_limit --format=csv,noheader,nounits -i 0 | tr -d ' ' | cut -d. -f1)"; [ -n "$d" ] && nvidia-smi -i 0 -pl "$d" >/dev/null 2>&1
+cd /root/fleet/out 2>/dev/null && for f in matrix.log ember.log prover.log prover-launch.log rows.jsonl; do [ -f "$f" ] && mv "$f" "$f.$(date +%s).old"; done
+echo "count=$(pgrep -c -f '^bash in/box-|^python3 -u /root/fleet/in/box-prover')+$(pgrep -c -x igneum-miner)+$(pgrep -c -f '^/opt/igneum/pkg/bin/igneum-worker-cuda')+$(pgrep -c -x sp1-gpu-server)+$(pgrep -c -f '^/opt/igneum-floor/(bin|target|prove)/.*igneum-prove-host') mem=$(nvidia-smi --query-gpu=memory.used --format=csv,noheader,nounits -i 0 | tr -d ' ')"
diff --git a/tools/fleet/fleet.py b/tools/fleet/fleet.py
index e5ecb6dc3..f04deb711 100755
--- a/tools/fleet/fleet.py
+++ b/tools/fleet/fleet.py
@@ -124,6 +124,9 @@ def run(script, labels, env=""):
     for iid, b in boxes(labels).items():
         if scp(b, [os.path.join(HERE, script)], "/root/fleet/in/") != 0: print(b["label"], "scp failed"); continue
         name = os.path.basename(script)
+        logname = {"box-matrix.sh": "matrix", "box-ember.sh": "ember", "box-prover.sh": "prover"}.get(name, name)
+        # rotate the stage's log before the start: an old run's end line must never be read as this run's (the 5070, 12:33Z)
+        ssh(b, f"cd /root/fleet/out 2>/dev/null && for f in {logname}.log {logname}-launch.log rows.jsonl; do [ -f $f ] && mv $f $f.$(date +%s).old; done; true", timeout=30)
         ssh(b, f"cd /root/fleet && chmod +x in/{name} && LABEL={b['label']} WALLET={b['wallet']} ARCHS={b['archs']} {env} setsid nohup in/{name} </dev/null >/dev/null 2>&1 & echo ok", timeout=60)
         print(b["label"], name, "started")