diff --git a/docs/port/HANDOFF.md b/docs/port/HANDOFF.md
index afe6bb40..879d9596 100644
--- a/docs/port/HANDOFF.md
+++ b/docs/port/HANDOFF.md
@@ -6449,3 +6449,32 @@ batching rule.
 Disc-wide: 970 elements, 91 name-stems, **958 (98.8 %) contain "btn"**. The only
 12 exceptions are `psselect_slot` and `psselect_slot_blank` — save-slot rows,
 which are menu items. The class is uniformly menu-row-shaped.
+
+## 🔴 Out-of-sample test failed 3 of 6 — withdraw the separation figure (`a61c191`)
+
+`docs/re/f6-out-of-sample-RESULT.md`. Predictions registered before the capture.
+
+| | predicted | fresh boot `f6c` | |
+|---|---|---|---|
+| leaf period ratio | 1.200 | **1.1753** | ✅ |
+| strip size ratio | 1.333 | **1.3009** | ✅ |
+| pulse / sweep loop | 0.100 | **0.0963** | ✅ |
+| pulse amplitude vs declared | ≤3 levels | **8.73** | 🔴 |
+| `ptcopyright` ramp ratio | 0.733 | **0.550** | 🔴 |
+| sweep leads plate | 0.138–0.141 | **0.0996** | 🔴 |
+
+✅ **Everything you build on is in the passing half.** `rate = 0.5` (leaf period
+ratio), the strip identities, the 120-unit pulse — all now 3-for-3 across
+captures, including one that had no hand in deriving them.
+
+🔴 **Withdraw the sweep→plate separation.** Three runs give **0.138, 0.141,
+0.0996**. My "agreeing to 2.2 %" was n=2, and two runs agreeing is not
+reproducibility. **The F6 conclusion is unaffected** — the glow still clearly
+precedes the plate with no input, and still coincides with Ⓐ — but there is no
+constant lead to author against. If you kept 0.14 anywhere as a check value,
+drop it.
+
+⚠️ And `check_labels.py`, which I offered you last iteration as the mechanism for
+capture-only labels, **fails its first independent test**: I validated it on the
+two captures that produced the labels. Two of four checks fire on a third. Treat
+it as unproven, not as a check.