diff --git a/docs/re/BACKLOG.md b/docs/re/BACKLOG.md
index 520034f..5a52dc4 100644
--- a/docs/re/BACKLOG.md
+++ b/docs/re/BACKLOG.md
@@ -454,9 +454,12 @@ search cannot find a *schedule*.
   90 s, and the new calibration added **two more 32 MB reads** — the 45 s freeze
   came right after it. 🔴 **The trim BROKE the witness** (17 candidates, `0/17` on
   every sample of a run with 13 losses) — reverted; two 32 MB reads once at
-  startup is the price of a working witness. ✅ **Instead, disabling the periodic
-  rescan removed the freeze**: first fully clean probed run, **zero stalls over
-  210 s** with 8 losses, ended by the turn timeout not a freeze (n=1). The rescan
+  startup is the price of a working witness. ✅🟡 **Instead, disabling the periodic
+  rescan LARGELY removes the freeze** — corrected from "removes" after more runs:
+  clean at 210/240/300 s but **frozen at 60 s** on a fourth. Tally — heavy probe
+  froze at 27/45/83/183/255 s; cheap+rescan at 183 s; cheap, no rescan: 3 of 4
+  clean past 200 s. A large probability improvement, **not a fix**. Practical
+  rule: run, check the witness, discard frozen runs (~3 in 4 usable). The rescan
   existed only to catch newly-allocated craft, which the roster work showed never
   happens.
 * ✅ **Run 16: the first TRUSTWORTHY negative.** Validated witness, no stall on
diff --git a/docs/re/guest-stalls.md b/docs/re/guest-stalls.md
index ec1a7ae..294c68c 100644
--- a/docs/re/guest-stalls.md
+++ b/docs/re/guest-stalls.md
@@ -376,3 +376,40 @@ clean no-probe control it points clearly at recurring heavy reads rather than at
 memory reading as such.
 
 The run ended on the turn's timeout, not a freeze.
+
+---
+
+# Freezes are stochastic, not eliminated (2026-08-24)
+
+The previous entry claimed disabling the periodic rescan "removes the freeze" on
+n = 1. With more runs that is too strong. The tally:
+
+| configuration | freeze onset |
+|---|---|
+| heavy probe | 27, 45, 83, 183, 255 s |
+| cheap probe + 90 s rescan | 183 s |
+| **cheap probe, no rescan** | clean 210 s, clean 240 s, clean 300 s, **frozen at 60 s** |
+
+Three of four runs now survive past 200 s where essentially none did before, so
+removing the rescan is a **real and large improvement** — but it is a change in
+probability, not a fix, and the wording is corrected accordingly.
+
+The witness handled the frozen run correctly: losses stop at t = 45 s, stalls are
+flagged from t = 60 s onward, no contradiction.
+
+## ✅ Calibration is no longer the bottleneck
+
+The candidate search was the last Python loop over 8 million words. Replaced with
+two numpy vector operations (`AGENT.md` notes numpy is installed, and its absence
+used to look like a logic bug). 7250 candidates found, and the startup cost is
+now small enough that **boot dominates**: ~190 s of title movie plus ~35 s to
+flight, against a 595 s turn cap, leaves roughly 350 s of observation.
+
+Boot is now the binding constraint on how much game time one turn can see, and
+it is the only remaining lever worth pulling if longer windows are needed.
+
+## Practical consequence
+
+Do not treat a single frozen run as evidence of anything. Run, check the witness,
+discard the frozen ones, and keep the clean runs — about three in four are usable
+now, which is workable where it previously was not.
diff --git a/tools/re-capture/wave7_probe.py b/tools/re-capture/wave7_probe.py
index c79179b..2504903 100755
--- a/tools/re-capture/wave7_probe.py
+++ b/tools/re-capture/wave7_probe.py
@@ -126,11 +126,18 @@ def main():
             n = min(1 << 24, hiw - pos); out += os.pread(fd, n, pos); pos += n
         return bytes(out)
     a = grab(); time.sleep(3.0); b = grab()
-    cands = []
-    for k in range(0, min(len(a), len(b)) - 3, 4):
-        va, vb = struct.unpack_from('>I', a, k)[0], struct.unpack_from('>I', b, k)[0]
-        if va < vb and 5 < (vb - va) / 3.0 < 200:
-            cands.append((lo + k, round((vb - va) / 3.0)))
+    # The candidate search was the last Python loop over 8 million words and cost
+    # most of the remaining startup time. numpy does the same comparison as two
+    # vector ops. (AGENT.md: numpy is installed; its absence used to look like a
+    # logic bug.)
+    import numpy as np
+    n4 = min(len(a), len(b)) // 4 * 4
+    A = np.frombuffer(a[:n4], dtype='>u4').astype(np.int64)
+    B = np.frombuffer(b[:n4], dtype='>u4').astype(np.int64)
+    d = B - A
+    rate = d / 3.0
+    idx = np.nonzero((d > 0) & (rate > 5) & (rate < 200))[0]
+    cands = [(lo + int(i) * 4, int(round(float(rate[i])))) for i in idx]
     # The MODAL cluster is not the frame counter. One run picked a modal rate of
     # 93/s, and only 11 of 31 of those advanced during active combat -- they are
     # subsystem counters that tick in bursts. timer_probe measured the frame-rate