diff --git a/tools_pc/framediff.py b/tools_pc/framediff.py
index c8372161..caba350f 100644
--- a/tools_pc/framediff.py
+++ b/tools_pc/framediff.py
@@ -51,7 +51,7 @@ threshold, 1 = at least one frame/region over threshold, 2 = usage / IO
 error (e.g. a golden frame with no candidate).
 
 Regenerate the golden set (after a deliberate visual change):
-  GE_PCDUMP="400-880:240" ./build-pc/ge007.x86_64.exe -level_09
+  GE_PCDUMP="640-1120:240" ./build-pc/ge007.x86_64.exe -level_09
   (with data/ge007.ini = only [Window] Width=640 Height=480; see
   tools_pc/golden/README.md)
   python tools_pc/framediff.py ppm --update
diff --git a/tools_pc/golden/README.md b/tools_pc/golden/README.md
index 37e8fb75..e2d71774 100644
--- a/tools_pc/golden/README.md
+++ b/tools_pc/golden/README.md
@@ -1,7 +1,7 @@
 # Golden baseline frames
 
 `frame_NNNNNN.png` — reference `GE_PCDUMP` captures for `framediff.py`
-(`-level_09` BUNKER1: end of the intro + settled gameplay). Compared structurally, not exactly — the port
+(`-level_09` BUNKER1: settled gameplay after the intro). Compared structurally, not exactly — the port
 is not frame-deterministic (see `docs/internals.md` "D117").
 
 Regenerate after a deliberate visual change:
@@ -13,7 +13,7 @@ printf '[Window]
 Width = 640
 Height = 480
 ' > data/ge007.ini
-GE_PCDUMP="400-880:240" ./build-pc/ge007.x86_64.exe -level_09
+GE_PCDUMP="640-1120:240" ./build-pc/ge007.x86_64.exe -level_09
 python tools_pc/framediff.py ppm --update    # then git rm any stale stems
 ```
 
@@ -30,7 +30,12 @@ python tools_pc/framediff.py ppm --update    # then git rm any stale stems
   it to 60 fps (1 tick/frame), so frame N now lands at half the sim time, and
   frames 200/320/440 fell in the intro cutscene, whose camera is
   nondeterministic (D117); two fresh captures disagreed at 440. The window
-  moved to `400-880:240`, the same sim ticks the original set sampled.
+  moved to `640-1120:240`, the same sim ticks the original set sampled.
   Two independent captures of it agree 3/3. The capture ini is now
   defaults-only, and `verify.sh` pins it the same way (it used to rewrite
   only Width/Height, which kept personal settings).
+- **2026-09-23 (later):** window moved again, `400-880:240` → `640-1120:240`.
+  Frame 400 is the last beat of the intro cutscene (Bond in the shaft) and
+  its camera is nondeterministic (D117): 1 in ~5 gate runs failed it (dmean
+  0.7 vs 20.7 run to run). All three frames are now settled gameplay; three
+  independent captures agree (worst dmean ≤ 6.8).
diff --git a/tools_pc/golden/frame_000400.png b/tools_pc/golden/frame_000400.png
deleted file mode 100644
index 06bc5c25..00000000
Binary files a/tools_pc/golden/frame_000400.png and /dev/null differ
diff --git a/tools_pc/golden/frame_000640.png b/tools_pc/golden/frame_000640.png
index a8c46e35..b89243e6 100644
Binary files a/tools_pc/golden/frame_000640.png and b/tools_pc/golden/frame_000640.png differ
diff --git a/tools_pc/golden/frame_000880.png b/tools_pc/golden/frame_000880.png
index 79040964..9a5dabf4 100644
Binary files a/tools_pc/golden/frame_000880.png and b/tools_pc/golden/frame_000880.png differ
diff --git a/tools_pc/golden/frame_001120.png b/tools_pc/golden/frame_001120.png
new file mode 100644
index 00000000..26fadafc
Binary files /dev/null and b/tools_pc/golden/frame_001120.png differ
diff --git a/tools_pc/verify.sh b/tools_pc/verify.sh
index 33eb86cb..71e044ac 100644
--- a/tools_pc/verify.sh
+++ b/tools_pc/verify.sh
@@ -18,7 +18,7 @@
 #
 # Options:
 #   --platform win|linux   default: autodetect via `uname -s`
-#   --dump lo-hi:step      GE_PCDUMP window (default: 400-880:240 single level,
+#   --dump lo-hi:step      GE_PCDUMP window (default: 640-1120:240 single level,
 #                          80-400:40 sweep — matches the existing golden set
 #                          and level_sweep.sh respectively)
 #   --script "..."         GE_INPUTSCRIPT passthrough (scripted input)
@@ -284,7 +284,7 @@ except Exception:
           # exit 1 = a real per-frame threshold fail (pixel/cell/hash delta).
           # exit 2 = a structural mismatch (missing/size) -- most commonly the
           # golden's frame stems don't line up with this run's --dump window
-          # (today's only golden, level_09, was captured at 400-880:240; a
+          # (today's only golden, level_09, was captured at 640-1120:240; a
           # sweep using a different stride has nothing to compare against
           # rather than a real regression). Only exit 1 is a verdict fail.
           case "$fexit" in
@@ -327,7 +327,7 @@ case "$MODE" in
     name="${resolved%%:*}"; num="${resolved##*:}"
     [ "${resolved%%:*}" = num ] && name=$(name_for_num "$num")
     pin_ini_640x480 2>/dev/null || true
-    verify_level "$name" "$num" "${DUMP:-400-880:240}"
+    verify_level "$name" "$num" "${DUMP:-640-1120:240}"
     ;;
 
   sweep)
@@ -375,7 +375,7 @@ case "$MODE" in
     [ -z "$resolved" ] && { echo "unknown level '$LEVEL_ARG'" >&2; exit 2; }
     name="${resolved%%:*}"; num="${resolved##*:}"
     pin_ini_640x480 2>/dev/null || true
-    run_one "$name" "$num" "${DUMP:-400-880:240}" 90
+    run_one "$name" "$num" "${DUMP:-640-1120:240}" 90
     if [ "$R_STATUS" != PASS ]; then
       emit_verdict "$name" "$R_STATUS" "$R_FRAMES" "" "$R_SYM"
       rm -rf "$CAPDIR"; exit 1