diff --git a/docs/re/f6-what-the-human-is-watching.md b/docs/re/f6-what-the-human-is-watching.md
new file mode 100644
index 00000000..a85b76de
--- /dev/null
+++ b/docs/re/f6-what-the-human-is-watching.md
@@ -0,0 +1,66 @@
+# F6 — the human is watching **their own Ⓐ press**, not either declared event
+
+**Question:** why does a person see the glow start exactly when the plate
+appears, when the declared data starts it far earlier?
+
+**What the human looks at:** boot, and press Ⓐ early — the glow and the
+`PRESS Ⓐ` plate come up together. Then boot again and press nothing: the glow
+arrives clearly first, well before the plate.
+
+**What this does NOT cover:** the `238…250` exit ramp (already settled as an
+exit), and anything about the menu.
+
+**Instrument:** ⟨capture⟩ ×4 — two no-input controls, two with a press.
+
+## The brief asked which of two declared events this is. It is neither.
+
+| | sweep pair enters | `ptcopyright` | separation |
+|---|---|---|---|
+| `f6b` no input | f577 | f611 | **34 frames** |
+| `f6` no input | f746 | f813 | **67 frames** |
+| `f5-early` Ⓐ during build-in | f438 | f436 | **2 frames** |
+| `f5` Ⓐ *after* the sweep already started | f427 | f456 | 29 frames |
+
+Phase-free, as a fraction of the sweep's own 600-unit loop: the controls give
+**34/600 = 0.057** and **67/1168 = 0.057** — agreeing to **1.2 %** across runs
+that differ 2× in frames — against **~0.003** when Ⓐ lands early.
+
+**Ⓐ collapses the gap by a factor of ~18.** It reveals the plate and restarts
+both sweep leaves at their own `t=0` in the same frame
+([`f5-verified-with-full-quad-reader.md`](f5-verified-with-full-quad-reader.md)),
+so the two events become simultaneous *because the player caused both*.
+
+## Why the report and the data both stand
+
+The human presses Ⓐ to get through the boot — that is what a player does — and
+at that instant the plate appears and the glow restarts together. Their
+observation is accurate. It is a description of **a boot with a press in it**,
+not of the declared timeline.
+
+The last row is the check that makes this an explanation rather than a story: in
+`f5` the press landed *after* the sweep had already started on its own, and the
+two are **29 frames apart**. Same input, opposite result, decided only by when
+the press falls.
+
+## 🔴 What the port must not do
+
+**Do not start the sweep when the plate appears.** That reproduces one boot and
+breaks the other: with no input the glow genuinely precedes the plate by 0.057 of
+a sweep loop. The correct behaviour is both of the pieces already delivered —
+
+* the declared parent gate, α0 until `t=70`, full at `t=100`; **and**
+* Ⓐ restarting both leaves at `t=0` as it advances the shared clock.
+
+Together those produce the human's observation *and* the no-input one, with
+nothing authored.
+
+## Refutation attempt
+
+I tried to find a case where the port's current behaviour (declared gate + snap
+restarting leaves) fails against these captures, and **could not**. All four runs
+are consistent with it. Recorded as an attempt that did not land.
+
+## Not settled
+
+* The per-cycle phase fit for the post-wrap decomposition.
+* Wrap-to-wrap sweep period: still one wrap per capture.