diff --git a/docs/re/f6-unit4-sweep-is-onscreen-early.md b/docs/re/f6-unit4-sweep-is-onscreen-early.md
new file mode 100644
index 00000000..ff48802c
--- /dev/null
+++ b/docs/re/f6-unit4-sweep-is-onscreen-early.md
@@ -0,0 +1,71 @@
+# F6 unit 4 — the sweep is **on screen 41 frames after the title starts**, long before the plate
+
+**Status: ✅ measured ⟨capture⟩ ×1.** 2026-09-02.
+
+**Question:** how many frames after the title starts drawing does the sweep first
+appear on screen?
+**Look at:** the frame its leading edge crosses into the viewport, minus the frame
+the title's first element draws.
+**Not covered:** alpha gating (unit 2, 🟡); the title's own clock origin.
+
+---
+
+## The measurement
+
+Same capture as units 2 and 3. Counts only — no clock conversion.
+
+```
+title's first drawn element   frame 706
+sweep first SUBMITTED         frame 746    (+40 frames)
+sweep first ON SCREEN         frame 747    (+41 frames)   NDC x span [−2.380, −0.990]
+```
+
+> **The sweep's leading edge is inside the viewport 41 frames after the title
+> begins drawing.**
+
+**For scale:** the plate reaches full alpha at declared `t = 236`. At the
+1 unit/present the splash work established, that is frame ≈ 942 — **195 frames
+after the sweep is already on screen.**
+
+⚠️ **That last comparison carries two assumptions and I am naming them rather
+than folding them in:** that the title's first drawn frame is its `t = 0`, and
+that its clock runs at 1 unit/present. The **41** is a raw frame count and carries
+neither.
+
+## 🔴 What this settles, and what it hands to unit 2
+
+**Position does not gate the sweep late.** It is geometrically visible from very
+early in the title, not at the plate. So if a human sees it start when the plate
+appears, **the gate cannot be where the quad is** — it has to be how opaque it is.
+
+That puts the whole of F6 on **unit 2**, which is 🟡: does the parent's alpha
+multiply into the leaf? The parent ramps `0→255` across `t = 70→100`, so:
+
+| if parent alpha multiplies in | the sweep is invisible until ≈ t 70 and dim to 100 — **a real early gate** |
+|---|---|
+| **if it does not** | the sweep is fully lit from the moment it enters at +41 frames — **and the human's report has no mechanism yet** |
+
+**Neither is established.** Unit 2's evidence was 14 sub-floor samples of 1754 and
+was rightly downgraded.
+
+📌 **But note the shape of the disagreement has narrowed usefully.** Before this
+unit, "starts when the plate appears" had two candidate mechanisms — position or
+alpha. Position is now excluded by measurement. One candidate left, and it is the
+one already under test.
+
+## Refutation attempt
+
+**Target:** my own unit 1 framing, that the human might be watching the leaf
+travel on screen rather than an alpha gate.
+
+**Result: REFUTED.** I wrote that a quad can be *"fully opaque and entirely
+off-screen"* and offered that as the likely explanation. It is not: the sweep is
+on screen at +41 frames, nowhere near the plate. **The travel explanation is
+dead**, and I am recording it because I proposed it two units ago and it steered
+the port's reading of their own renderer.
+
+## Reach
+
+⟨capture⟩ ×1, one boot, `pteff03` only. The `+41` is a frame count between two
+events in one stream and needs no clock. The comparison to the plate does need
+one, and is labelled. `pteff03a` does not appear in this capture.