diff --git a/CHANGELOG.md b/CHANGELOG.md
index 3c2d2cb77..6c560005d 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -10,6 +10,10 @@ installable release; see the roadmap in [README.md](README.md).
 
 ## [Unreleased]
 
+### Performance
+
+- **`hrr.DEFAULT_DIM` flips 2048 → 512** ([#538](https://github.com/robotrocketscience/aelfrice/issues/538)). Aligns the HRR default dimensionality with the lab campaign R4 verdict ("adopt dim=512 default; dim=2048 escape hatch") that produced the spec at `docs/hrr_structural_query_lane.md`. ~4× memory footprint reduction for the structural-query lane: ~200 MB at N=50k vs the prior ~800 MB. Capacity per Plate's bound stays well above realistic per-belief role multiplicity (~57 retrievable bound pairs at dim=512 vs ~5 typical outgoing edges per belief). Set `dim=2048` explicitly via the `HRRStructIndex(dim=...)` kwarg for high-multiplicity corpora.
+
 ### Changed
 
 - **`use_heat_kernel` and `use_hrr_structural` flip to default-on** ([#154](https://github.com/robotrocketscience/aelfrice/issues/154)). Both retrieval-lane flags shipped opt-in earlier (heat-kernel #150 in v1.6, HRR structural-query lane #152 in v1.7) pending the calibrated reproducibility-harness gate. That gate cleared at 11/11 via [PR #489](https://github.com/robotrocketscience/aelfrice/pull/489) closing #437, so the #154 composition tracker flips the defaults per its status banner. Precedence (env > kwarg > TOML > default) is unchanged; only the default values flip `False` → `True`. Reversible via `[retrieval] use_heat_kernel = false` / `use_hrr_structural = false` in `.aelfrice.toml` (or `AELFRICE_HEAT_KERNEL=0` / `AELFRICE_HRR_STRUCTURAL=0`) for parity with the v2.0.x ranking.
diff --git a/src/aelfrice/hrr.py b/src/aelfrice/hrr.py
index 0cee8ba4b..b9ab19c80 100644
--- a/src/aelfrice/hrr.py
+++ b/src/aelfrice/hrr.py
@@ -28,10 +28,13 @@
 Vector = npt.NDArray[np.float64]
 
 # Default dimensionality. Capacity per Plate's analysis is roughly
-# dim/9 retrievable bound pairs at signal-to-noise threshold; 2048
-# accommodates ~227 distinct bindings, well above aelfrice's typical
-# ~5 outgoing edges per belief.
-DEFAULT_DIM: Final[int] = 2048
+# dim/9 retrievable bound pairs at signal-to-noise threshold; 512
+# accommodates ~57 distinct bindings, well above aelfrice's typical
+# ~5 outgoing edges per belief. dim=2048 is the escape hatch for
+# high-multiplicity corpora where K approaches the lower bound.
+# See #538 / lab campaign R4 verdict (full-corpus latency budget at
+# N=100k holds at dim=512; dim=2048 does not).
+DEFAULT_DIM: Final[int] = 512
 
 
 # ---------------------------------------------------------------------------
diff --git a/src/aelfrice/hrr_index.py b/src/aelfrice/hrr_index.py
index 056ca21af..c48a06ab3 100644
--- a/src/aelfrice/hrr_index.py
+++ b/src/aelfrice/hrr_index.py
@@ -11,8 +11,8 @@
 ``bind(role[KIND], id_vec[target])`` and ranks beliefs by the
 inner product against ``struct[b]``. Beliefs whose structure
 contains exactly that bound term score high; orthogonal noise from
-other bindings is ``~1/sqrt(dim)`` per term, so at ``dim=2048`` the
-top-K is dominated by true structural matches.
+other bindings is ``~1/sqrt(dim)`` per term, so at the default
+``dim=512`` the top-K is dominated by true structural matches.
 
 Parallel to the textual lane (BM25F + heat kernel), not a competitor.
 A query parser (:func:`parse_structural_marker`) routes structural
@@ -106,7 +106,9 @@ class HRRStructIndex:
     Build is offline (walks every edge once); query is one matvec
     against the ``(N, dim)`` struct matrix plus one bind. Storage
     cost is dominated by the struct matrix at ``8 * N * dim`` bytes;
-    at ``N=50k, dim=2048`` that is ~800 MB, fitting the AC8 budget.
+    at ``N=50k`` and the default ``dim=512`` that is ~200 MB
+    (~800 MB at the ``dim=2048`` escape-hatch value), both within
+    the AC8 budget.
 
     Determinism: ``random_vector`` draws are reproducible from
     ``np.random.default_rng(seed)``. Two builds against the same