diff --git a/docs/re/entity-position-anchor-refuted.md b/docs/re/entity-position-anchor-refuted.md
new file mode 100644
index 0000000..16781cd
--- /dev/null
+++ b/docs/re/entity-position-anchor-refuted.md
@@ -0,0 +1,86 @@
+# ❌ Two of my own claims, refuted — and a live-memory trap behind both
+
+Measured 2026-08-26, Canary on the retail disc, save slot 01, Stage 01, **in
+flight with a live HUD** (`TIME 00:28.42`, speed 350, target on radar, dialogue
+mid-render — the game was unambiguously running throughout).
+
+## ❌ Refuted: `position = instance − 0x12c`
+
+The corpus anchors the entity layout on **position**:
+[`autopilot-memory-driven.md`](autopilot-memory-driven.md) gives the definition
+pointer at `position + 0x130`, orientation at `position − 0x70`, hull at
+`position + 0x154`. `gworld.py`'s `name_of()` reads the definition pointer at
+`instance + 4`. If those name the same field, `instance + 4 = position + 0x130`,
+so `position = instance − 0x12c`.
+
+They do not. Read at `instance − 0x12c` for all **42** named instances:
+
+```
+UN_f001_TCAF_DeltaSaber_T_Player   (     0.0,      0.9,      0.1)   moved 0.00
+UN_e106_ADAN_Destroyer             (   -48.9,      0.0,      0.0)   moved 0.00
+UN_S01_Asteroid_cmesh_01a          (2.07e11,      0.0,      0.0)   moved 0.00
+...
+moved in 0.8 s : 0/42
+```
+
+Almost every read is `(0, 0, 0)`, three are garbage, and **nothing moves**. The
+vtable-bearing object and the position block are two different structures that
+merely name the same units. The prediction was clean and it was wrong.
+
+## ❌ Withdrawn: "`entities2.py`'s VA window is aimed at the wrong region"
+
+[Last iteration](world-unit-attempt-2026-08-26.md) I recorded `ENT_VA_LO/HI =
+0xBD000000–0xBE000000` as a defect, on the grounds that the live instances sit at
+`0xBC384CE0–0xBC9BAC20` and the window "covers the definitions instead".
+
+That reasoning was wrong. It assumed the thing the window *should* contain is the
+vtable-bearing instance — but `entities2.py` does not look for those. It hunts
+**moving position triples** and checks `+0x130` for a definition pointer, exactly
+the documented anchor. And in the run where I called it broken it had *worked*
+moments earlier: **6914 movers, `+0x130` voted 92×**, then 6634 and 138× on a
+re-run. I built a defect report on the one sample in between that returned zero —
+the same single-sample error this corpus keeps logging.
+
+The window may still be imperfect (it is hardcoded, and the heap moves). But it
+is not "aimed at the definitions", and nothing here shows it needs changing.
+
+## 🟡→❌ The `+0x29d0` position candidate is not trustworthy either
+
+It was read through the same region that is shown below not to be updating, so
+what it captured is load-time data, not a live position. Withdrawn pending a
+re-measurement through a region that demonstrably changes.
+
+## ✅ The trap: most of guest memory is not changing while the game runs
+
+Sampling **64 spots of 64 KB** spread across the mapping's **505 extents
+(357 MB)**, twice, 2.5 s apart, with the mission visibly running:
+
+```
+CHANGING: 5 / 64
+   0x007006f000   0x00704db000   0x0070557b000   0x00705df000   0x0080240000
+```
+
+Everything else — including the megabyte containing all 42 live instances
+(`0x11c37d780`), and that region's primary-VA counterpart at `0x00bc300000` — is
+**byte-identical over seconds**. So:
+
+* The mapping is live. Five regions prove it, so "the shm view is stale" as a
+  blanket claim is refuted.
+* The entity-object region genuinely does not change on a seconds timescale.
+* ❔ **Not determined**: whether that is because those records are static
+  metadata and the moving state lives elsewhere, or because the guest physical
+  page is aliased at several file offsets and writes land in a different one.
+  Both fit the evidence. `gmem.primary_va(0x11c384ce0) = 0xbc384ce0` shows the
+  alias relationship exists; it does not show which copy the guest writes.
+
+**Consequence for anyone reading guest memory here:** verify the specific region
+you are sampling actually changes before drawing conclusions from it. A scan that
+finds correct, name-resolving records can still be reading a snapshot.
+
+## Still open
+
+* Where the live position blocks are. The five changing offsets above are the
+  place to start, but `moving()` takes **VAs** and converts with
+  `gmem.va_to_off`, while those five are **file offsets** — they are not the same
+  number and must be converted before they can be scanned.
+* Whether `entities2.py`'s window needs any change at all.