diff --git a/docs/re/f6-unit9-sweep-period-and-onset.md b/docs/re/f6-unit9-sweep-period-and-onset.md
new file mode 100644
index 00000000..090eb314
--- /dev/null
+++ b/docs/re/f6-unit9-sweep-period-and-onset.md
@@ -0,0 +1,76 @@
+# F6 unit 9 — the sweep's loop period, and when it starts
+
+**Question (one sentence):** does the title's sweeping glow start when the plate
+appears, and how long is one pass?
+
+**What the human looks at:** boot the title, watch the white sweep. Pass = it
+first enters the screen shortly *before* the "PRESS Ⓐ" plate fades in — under a
+second before, not at the very start of the title.
+
+**What this does NOT cover:** which declared keyframe drives the sweep, whether
+the artwork and the plate share one clock (that is F4), and F5 entirely.
+
+**Instrument:** ⟨capture⟩ — two independent Canary runs, `f6` and `f6b`.
+
+## ⚠️ First, the thing that invalidates frame counts across captures
+
+The same animation took **1168 frames** in one run and **600** in the other —
+**1.947×** apart. Every other interval in those captures scales by the same
+factor (the baseline below: **1.953×**). So the runs differ in present rate by
+~1.95×, and **any timing quoted in captured frames is specific to its run.**
+
+This is why the numbers below are all ratios. It also means a frames-based rate
+carried from one capture to another is wrong by up to 2×.
+
+## Measured (phase-free, reproduced across both runs)
+
+Baseline interval = from the `x=-0.740` element appearing to the plate
+appearing. Chosen because it is long; a short alpha ramp quantises to ±1 frame
+and gave an 8% spread where this gives 0.35%.
+
+| | f6 | f6b | agreement |
+|---|---|---|---|
+| sweep loop period | 1168 fr | 600 fr | — (not comparable) |
+| baseline | 84 fr | 43 fr | — |
+| **period / baseline** | **13.905** | **13.953** | **0.35%** |
+| **sweep starts before plate** | **0.798** | **0.791** | **0.9%** |
+
+The cycle boundary is unambiguous: the sweep enters at `x=-1.540` and wraps from
+`x=+1.840` back to `x=-1.540`, so first-appearance→wrap is one whole period.
+
+Calibrating the baseline with the plate's declared 12-unit ramp (238→250) puts
+it at ~48–52 title units, so in title units: **period ≈ 670–720**, **the sweep
+leads the plate by ≈ 40**.
+
+## What this says about F6
+
+**The sweep does not start at title `t=0`, and it does not start exactly with the
+plate.** It starts a short, reproducible interval *before* the plate — ~0.8 of
+the baseline, ≈40 title units, well under a second.
+
+So the human's report is approximately right and the port's behaviour is not: at
+`t=0` the port starts it roughly 200 units early. The human could not have seen a
+40-unit lead as separate from the plate, and did not claim to.
+
+## 🟡 A conflict I did not smooth
+
+Unit 8 measured the title clock at **1.0000 units/frame** in `f6` from
+`ptcopyright`'s 22-unit ramp taking 22 frames. In the *same capture*, the plate's
+declared 12-unit ramp takes 21 frames — **0.571 units/frame**, 1.75× apart.
+
+One of three things is true: `ptcopyright`'s ramp is not 22 units, the plate's is
+not 12, or **the two elements are not on one clock** — which is exactly the
+premise (`clock: "shared"`) the port's `flow.json` is built on. Not resolved
+here; it needs F4.
+
+Because of this, the title-unit conversions above are 🟡. The **ratios are ✅** —
+they need no clock at all.
+
+## Not settled
+
+* Still **one wrap per capture**, so the period is start→wrap, not wrap→wrap. It
+  is a full cycle by the x-boundary argument, but a second wrap would be better.
+  The title exits before one arrives; `f6b` ran 504 frames past the wrap and the
+  next was ~100 short.
+* Which declared keyframe the sweep's leaf hangs off.
+* F5, untouched.
diff --git a/tools/re-capture/title_sweep_probe.sh b/tools/re-capture/title_sweep_probe.sh
index 85d99db9..931387d8 100755
--- a/tools/re-capture/title_sweep_probe.sh
+++ b/tools/re-capture/title_sweep_probe.sh
@@ -25,19 +25,21 @@ win="$(xdotool search --name "Xenia-canary" | tail -1)"
 xdotool windowactivate "$win" 2>/dev/null; xdotool key --window "$win" F10; xdotool key F10
 echo "armed at ${SECONDS}s"
 
+sleep 12; pad "press=A"; sleep 0.4; pad ""; echo "A (skip video) at ${SECONDS}s"
+
 # ⚠️ ARMING IS NOT CONFIRMED BY SENDING THE KEY. Twice now a probe has printed
 # "armed", pressed on, and written no draw log at all -- once because the window
 # lookup failed, once with the window found and the key sent. A probe that cannot
 # confirm its own instrument is recording is a probe whose negatives mean nothing.
 # So: wait for the log to exist AND grow, and abort loudly if it does not.
-for _ in $(seq 1 20); do
+for _ in $(seq 1 45); do
   sz=$(stat -c %s "$OUT"/xenia_re_ui_draws_*.log 2>/dev/null | head -1 || echo 0)
   [ "${sz:-0}" -gt 0 ] && break
   sleep 1
 done
-[ "${sz:-0}" -gt 0 ] || { echo "FATAL: armed but no draw log after 20s -- not recording"; pkill -x xenia_canary; exit 1; }
+[ "${sz:-0}" -gt 0 ] || { echo "FATAL: armed but no draw log after 45s -- not recording"; pkill -x xenia_canary; exit 1; }
 echo "logging confirmed at ${SECONDS}s (${sz} bytes)"
-sleep 12; pad "press=A"; sleep 0.4; pad ""; echo "A (skip video) at ${SECONDS}s"
-sleep 150         # let the title build in and settle, no further input
+
+sleep 260         # long enough for TWO leaf wraps, not one
 pkill -x xenia_canary
 echo "done at ${SECONDS}s"; ls -la "$OUT"/*.log 2>/dev/null