diff --git a/TODO.md b/TODO.md
index 51001bb..a06bfe3 100644
--- a/TODO.md
+++ b/TODO.md
@@ -878,3 +878,49 @@ and that nothing catches it.
 The content programme this opens up — what a single cohesive California is still
 missing, and in what order — is **`CALIFORNIA.md`**, kept separate because none
 of it needs a new mechanism and all of it can be worked in parallel.
+
+## Los Angeles at 7.7 km is over the triangle cap, and no budget cell stands there
+
+Found by the review of the culling work, 2026-08-24, and recorded here because
+the commit record for that work names a *different* LA pose — the 110 km one,
+which after the terrain split measures 310,841 and is comfortably under. The pose
+that is over the cap is not in any commit message, and this is the correction.
+
+Measured with `scripts/cost-at.mjs` on the merged board at 34.05, -118.24:
+
+| stand-off | before culling | after culling | cap |
+| --- | --- | --- | --- |
+| 110 km | 341,803 | 310,841 | 400,000 |
+| **7.7 km** | **773,282** | **604,332** | **400,000** |
+| 2.5 km | 761,648 | 335,186 | 400,000 |
+
+So the culling work **closed the 2.5 km breach and reduced the 7.7 km one by
+22%**, and 7.7 km is still 51% over. It is a pre-existing defect that this work
+improved rather than one it introduced — but it is live, it is reachable from the
+board's own chapters, and it was never measured because `california-one` poses at
+`california-overview` and nothing else does.
+
+**Frustum culling cannot fix it.** At that stand-off 40,025 detail lots are
+genuinely inside the frustum. The Southland is a basin ninety kilometres across
+and at 7.7 km you are looking down the length of it; there is no camera geometry
+that makes those buildings not on screen. What is wrong is the *lot size*: 160 m
+is chosen for a metro seen from its own stand-off, and at 7.7 km over the basin
+the board is drawing a hundred kilometres of city at that density.
+
+The fix is a lot-size level of detail — the same mechanism the near-field 40 m
+tier needs, pointing the other way. Two shapes, and the choice needs measuring
+rather than arguing:
+
+- **A stride.** Pack every Nth instance of a range, N from the stand-off. One
+  copy loop, no new geometry, no new draw call — the store already holds every
+  instance. The risk is shimmer: instance order within a range is lattice order,
+  so a stride is a spatially regular thinning and regular thinning of a grid is
+  exactly what aliases.
+- **A second, coarser instance set per district**, built at 480 m and swapped in
+  by stand-off. No shimmer, and it costs build time and memory for lots that are
+  usually not drawn.
+
+**Two things to do first, in this order.** Add a budget cell that stands at that
+pose — `no budget cell stands there` is the actual defect, and a cap nobody
+measures is not a cap. Expect it to fail on the day it lands; that is the point,
+and it must not be answered by raising a number. Then fix it.