diff --git a/CHANGELOG.md b/CHANGELOG.md
index e4b1fb84..4bff211d 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -24,6 +24,10 @@ installable release; see the roadmap in [README.md](README.md).
 
 - **Research-agent dispatch wonder surface** ([#551](https://github.com/robotrocketscience/aelfrice/issues/551), umbrella [#542](https://github.com/robotrocketscience/aelfrice/issues/542) track E). New `aelfrice.wonder.dispatch` module exposes `analyze_gaps()` and `generate_research_axes()`. `analyze_gaps(store, query, ...)` returns a `GapAnalysis` dataclass with known beliefs, high-uncertainty beliefs (normalized 0..1 variance proxy > 0.7), unresolved CONTRADICTS pairs (CONTRADICTS minus matching SUPERSEDES), query-term coverage, and named gaps. `generate_research_axes(gap, agent_count)` produces 2–6 orthogonal `ResearchAxis` records: always-on `domain_research` + `internal_gap_analysis`, conditional `contradiction_resolution` / `uncertainty_deep_dive` / `coverage_extension`. Surfaced via the new `aelf_wonder` MCP tool (13th tool) and `aelf wonder --axes QUERY` CLI flag — both return the same JSON shape for downstream skill-layer consumption (E4, separate sub-issue).
 
+### Changed
+
+- **`use_intentional_clustering` flips to default-on** ([#436](https://github.com/robotrocketscience/aelfrice/issues/436)). Substrate landed in v2.0/v2.1 ([PR #496](https://github.com/robotrocketscience/aelfrice/pull/496) module + corpus, [PR #498](https://github.com/robotrocketscience/aelfrice/pull/498) `retrieve_v2` wiring, [PR #504](https://github.com/robotrocketscience/aelfrice/pull/504) bench-gate scorer). Bench gate cleared on the production multi-store sweep (#436 R6: 60/60 PASS, p99 0.328ms — ~15-30× margin under the 5ms A4 latency budget). Per the `resolve_use_intentional_clustering()` docstring contract ("the bench gate flips the default after lab-side benchmark evidence clears"), the default is unblocked. Precedence (env > kwarg > TOML > default) is unchanged; only the default value flips `False` → `True`. Reversible via `[retrieval] use_intentional_clustering = false` in `.aelfrice.toml` (or `AELFRICE_INTENTIONAL_CLUSTERING=0`) for v2.0.x parity. The lab-finding that raising `DEFAULT_CLUSTER_EDGE_FLOOR` from 0.4 → 0.6 triples uplift (by excluding `EDGE_CITES` inter-cluster edges) is queued as a separate follow-up so this change stays minimal and reviewable.
+
 ### Documentation
 
 - **README documents default-on auto-capture** ([#557](https://github.com/robotrocketscience/aelfrice/issues/557)). Adds a passive-capture row to the `What it remembers` table and a `Passive capture` bullet to `What you get for free` so the headline narrative reflects the v2.1 default flip (#529). Pre-v2.1 the table listed only manual inputs, so a model reading the README inferred aelfrice was for user-locked rules only — the symptom in #557.
diff --git a/src/aelfrice/retrieval.py b/src/aelfrice/retrieval.py
index dd89a506..090e5cf2 100644
--- a/src/aelfrice/retrieval.py
+++ b/src/aelfrice/retrieval.py
@@ -908,9 +908,10 @@ def resolve_use_intentional_clustering(
       1. AELFRICE_INTENTIONAL_CLUSTERING env var (truthy / falsy normalised).
       2. Explicit `explicit` kwarg from the caller.
       3. `[retrieval] use_intentional_clustering` in `.aelfrice.toml`.
-      4. Default: False — ships behind the flag at v2.0.0; the bench gate
-         (A2 in docs/feature-intentional-clustering.md) flips the default
-         after lab-side benchmark evidence clears.
+      4. Default: True — flipped from False after the A4 latency bench
+         gate cleared on the multi-store production sweep (#436 R6, 60/60
+         PASS at p99 0.328ms ~ 15-30x margin under the 5ms budget). See
+         docs/feature-intentional-clustering.md A2 + A4.
     """
     env = _env_intentional_clustering_override()
     if env is not None:
@@ -920,7 +921,7 @@ def resolve_use_intentional_clustering(
     toml_value = _read_toml_flag_for(INTENTIONAL_CLUSTERING_FLAG, start)
     if toml_value is not None:
         return toml_value
-    return False
+    return True
 
 
 def resolve_use_vocab_bridge(
diff --git a/tests/test_clustering_integration.py b/tests/test_clustering_integration.py
index 95a670b8..a82abbf4 100644
--- a/tests/test_clustering_integration.py
+++ b/tests/test_clustering_integration.py
@@ -70,8 +70,8 @@ def _isolated_cwd(monkeypatch: pytest.MonkeyPatch, tmp_path: Path) -> Path:
 # --- Flag resolution ---------------------------------------------------
 
 
-def test_default_is_off(_no_env_override: None, _isolated_cwd: Path) -> None:
-    assert resolve_use_intentional_clustering() is False
+def test_default_is_on(_no_env_override: None, _isolated_cwd: Path) -> None:
+    assert resolve_use_intentional_clustering() is True
 
 
 def test_explicit_kwarg_overrides_default(
@@ -94,8 +94,8 @@ def test_env_garbage_falls_through(
     monkeypatch: pytest.MonkeyPatch, _isolated_cwd: Path
 ) -> None:
     monkeypatch.setenv(ENV_INTENTIONAL_CLUSTERING, "maybe")
-    assert resolve_use_intentional_clustering() is False
-    assert resolve_use_intentional_clustering(True) is True
+    assert resolve_use_intentional_clustering() is True
+    assert resolve_use_intentional_clustering(False) is False
 
 
 def test_toml_resolves_when_kwarg_and_env_unset(
@@ -146,27 +146,30 @@ def _populate_clustered_store() -> MemoryStore:
     return s
 
 
-def test_default_call_byte_identical_to_explicit_off(
+def test_default_call_byte_identical_to_explicit_on(
     _no_env_override: None, _isolated_cwd: Path
 ) -> None:
-    """Default and explicit-OFF must agree on the merged belief id list.
+    """Default and explicit-ON must agree on the merged belief id list.
 
-    This is the OFF-byte-identity invariant that makes the wiring safe to
-    land default-OFF — pre-#436-Phase-2 callers see no behavior change."""
+    Post-#436 default-flip: callers that don't pass the kwarg get the
+    same selection as callers that pass `use_intentional_clustering=True`.
+    Operators wanting v2.0.x parity opt out via
+    `[retrieval] use_intentional_clustering = false` or
+    `AELFRICE_INTENTIONAL_CLUSTERING=0`."""
     s = _populate_clustered_store()
-    explicit_off = retrieve_v2(
+    explicit_on = retrieve_v2(
         s, "deploy sqlite",
         budget=2400,
         use_entity_index=False,
-        use_intentional_clustering=False,
+        use_intentional_clustering=True,
     )
-    default_off = retrieve_v2(
+    default_on = retrieve_v2(
         s, "deploy sqlite",
         budget=2400,
         use_entity_index=False,
     )
-    assert [b.id for b in explicit_off.beliefs] \
-        == [b.id for b in default_off.beliefs]
+    assert [b.id for b in explicit_on.beliefs] \
+        == [b.id for b in default_on.beliefs]
 
 
 def test_clustering_changes_selection_at_tight_budget(
diff --git a/tests/test_compression_integration.py b/tests/test_compression_integration.py
index 373c3f6c..8441df55 100644
--- a/tests/test_compression_integration.py
+++ b/tests/test_compression_integration.py
@@ -153,7 +153,13 @@ def test_compressed_beliefs_populated_when_flag_on(
     _no_env_override: None, _isolated_cwd: Path
 ) -> None:
     s = _populate_store()
-    result = retrieve_v2(s, "sqlite system", use_type_aware_compression=True)
+    # use_intentional_clustering=False explicit: post-#436 default-flip,
+    # the clustering flag is default-on and would mutex with compression.
+    result = retrieve_v2(
+        s, "sqlite system",
+        use_type_aware_compression=True,
+        use_intentional_clustering=False,
+    )
     assert len(result.compressed_beliefs) == len(result.beliefs)
     # Same order, same belief ids.
     for b, cb in zip(result.beliefs, result.compressed_beliefs, strict=True):
@@ -164,7 +170,11 @@ def test_compression_strategy_dispatches_by_retention_class(
     _no_env_override: None, _isolated_cwd: Path
 ) -> None:
     s = _populate_store()
-    result = retrieve_v2(s, "sqlite system", use_type_aware_compression=True)
+    result = retrieve_v2(
+        s, "sqlite system",
+        use_type_aware_compression=True,
+        use_intentional_clustering=False,
+    )
     by_id = {cb.belief.id: cb for cb in result.compressed_beliefs}
     assert by_id["F1"].strategy == STRATEGY_VERBATIM
     assert by_id["S1"].strategy == STRATEGY_HEADLINE
@@ -177,6 +187,13 @@ def test_env_var_alone_enables_compression(
     monkeypatch: pytest.MonkeyPatch, _isolated_cwd: Path
 ) -> None:
     monkeypatch.setenv(ENV_TYPE_AWARE_COMPRESSION, "1")
+    # Post-#436 default-flip, AELFRICE_INTENTIONAL_CLUSTERING must also
+    # be disabled in this scope to satisfy the v2.0.0 mutex (the cluster
+    # pack accounts in raw tokens; composing it with compressed cost is
+    # tracked as a v2.x follow-up). The test still meaningfully exercises
+    # "env var alone enables compression" — it just makes the clustering
+    # env-disable explicit instead of relying on the (now-flipped) default.
+    monkeypatch.setenv("AELFRICE_INTENTIONAL_CLUSTERING", "0")
     s = _populate_store()
     result = retrieve_v2(s, "sqlite system")  # no explicit kwarg
     assert len(result.compressed_beliefs) == len(result.beliefs)
@@ -228,6 +245,7 @@ def test_pack_widens_when_flag_on(
         budget=80,
         use_entity_index=False,
         use_type_aware_compression=True,
+        use_intentional_clustering=False,
     )
     assert len(on.beliefs) > len(off.beliefs)
 
@@ -278,6 +296,7 @@ def test_pack_locked_unchanged_when_flag_on(
         s, "sqlite",
         use_entity_index=False,
         use_type_aware_compression=True,
+        use_intentional_clustering=False,
     )
     assert [b.id for b in off.beliefs] == [b.id for b in on.beliefs]
     # Locked render is verbatim under compression.