diff --git a/THREADS.md b/THREADS.md
index 73164c6..95ce389 100644
--- a/THREADS.md
+++ b/THREADS.md
@@ -9770,3 +9770,27 @@ anchors are your GLB), but the tooling is now waiting, not TODO.
     with one variable left in it instead of two.
     **AND THE SAME WIRING IS WHAT THE GLASSHOUSE TIER WILL WANT** for two-glasshouse yards, which is
     why I did not invent a second mechanism for it.
+
+[E] 2026-07-25 — ✏️ **CORRECTION TO MY OWN ENTRY, SAME SESSION: I overwrote the field I told you to
+    trust.** The gate-4 entry above says *"the shipped GLBs were exported by Blender 5.1.2;
+    `asset_report.json` records it in its `blender` field — check it before you build."* **That
+    instruction is now wrong, and I made it wrong.** My run regenerated the report, so the committed
+    `asset_report.json` reads **`"blender": "5.0.1"`** — this box — while 37 of the 39 GLBs beside it
+    are still the **5.1.2** exports I deliberately reverted. Read literally, the field now tells the
+    next session "you are on the right Blender" at the exact moment they are not, which is the same
+    shape as every trap this sprint went hunting: a source of truth that quietly starts answering a
+    different question.
+    **What is actually true of the committed tree:** `hail_net_frame_01_v1.glb` and
+    `hail_net_frame_01_wrecked_v1.glb` were built by **5.0.1**; the other 37 GLBs and all 8 textures
+    are **5.1.2** artifacts, untouched. The report's per-asset facts (dims, tris, node names, baked
+    extras) are correct for every one of the 39 either way — that is measured, not assumed, and it is
+    why reverting was safe.
+    **So `asset_report.json:blender` names THE LAST MACHINE THAT RAN THE FACTORY, not the machine
+    that built the GLBs you are looking at.** Those are the same thing only after a churn-free run.
+    The honest check is the one I actually used: build, then `git status` — **if files you did not
+    author come back modified, you are on the wrong Blender**, and the header of
+    `build_yard_assets.py` tells you what the differences mean. I am leaving the generated file alone
+    rather than hand-editing it (it says "do NOT hand-edit" for good reasons, and a doctored report is
+    worse than a mislabelled one); the fix is this paragraph, and the mixed-exporter state going into
+    the integrator's eyes on purpose. If someone runs the full factory on a 5.1.2 box, the churn will
+    be **my two files only** and the tree becomes uniform again — that is the cheap moment to do it.