diff --git a/CHANGELOG.md b/CHANGELOG.md
index 788d4e53..ae19b6d9 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -9,6 +9,23 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ## [Unreleased]
 
+## [0.2.7.post1] - 2026-07-03
+
+### Fixed
+
+- **Reassessment: stay serial on low-core hosts** (`utils/reassessment.py`). The
+  parallel per-action reassessment (0.2.6) clones the whole network per worker
+  (`save`/`load_from_binary_buffer` + `SimulationEnvironment` + observation
+  build) — a fixed overhead the serial path never pays, only amortized above
+  ~3-4 usable cores. On a **2-vCPU host** (e.g. a small HuggingFace Space) the
+  two clones + GIL contention on the Python-side observation build made the
+  "parallel" path **slower than serial** (observed ~42 s vs ~9 s on a many-core
+  Mac for the same 15-action pan-European case). Parallel reassessment is now
+  gated behind `min(10, cores, n_actions) >= 4` (env-tunable via
+  `EXPERT_OP4GRID_MIN_PARALLEL_REASSESS_WORKERS`; floored at 2). 2-vCPU
+  deployments fall back to the faster serial path; ≥4-core hosts keep the
+  speed-up. Behaviour and output are otherwise unchanged.
+
 ## [0.2.7] - 2026-07-03
 
 ### Changed
diff --git a/CLAUDE.md b/CLAUDE.md
index e09439b3..593e94ee 100644
--- a/CLAUDE.md
+++ b/CLAUDE.md
@@ -426,7 +426,12 @@ pytest tests/test_ActionClassifier.py::test_specific  # Single test
 
 ## Current Development Status
 
-**Current version**: `0.2.7` (see `CHANGELOG.md` for full history)
+**Current version**: `0.2.7.post1` (see `CHANGELOG.md` for full history)
+
+> **v0.2.7.post1**: reassessment stays serial on low-core hosts — the parallel
+> path's per-worker network clone is only amortized above ~4 cores, so a 2-vCPU
+> Space now uses the faster serial path (gate:
+> `EXPERT_OP4GRID_MIN_PARALLEL_REASSESS_WORKERS`, default 4).
 
 > **v0.2.7 highlights** (deep revisions R1 + R2 from the 2026-07 review): typed pipeline
 > spine — `AnalysisContext` / `AnalysisResult` dataclasses replace the ~41-key context dict
diff --git a/expert_op4grid_recommender/__init__.py b/expert_op4grid_recommender/__init__.py
index b933372f..a129a62d 100644
--- a/expert_op4grid_recommender/__init__.py
+++ b/expert_op4grid_recommender/__init__.py
@@ -15,7 +15,7 @@
 
 import logging
 
-__version__ = "0.2.7"
+__version__ = "0.2.7.post1"
 
 _logger = logging.getLogger(__name__)
 
diff --git a/expert_op4grid_recommender/utils/reassessment.py b/expert_op4grid_recommender/utils/reassessment.py
index 12ff564d..3c533661 100644
--- a/expert_op4grid_recommender/utils/reassessment.py
+++ b/expert_op4grid_recommender/utils/reassessment.py
@@ -34,6 +34,47 @@
 #: Hard cap on reassessment worker threads (see ``_reassessment_worker_count``).
 _MAX_REASSESSMENT_WORKERS = 10
 
+#: Default minimum worker count below which the parallel reassessment is NOT
+#: worth it. The parallel path clones the whole network per worker
+#: (``save``/``load_from_binary_buffer`` + ``SimulationEnvironment`` + obs
+#: build) — a fixed overhead the serial path never pays. That overhead is only
+#: amortized above ~3-4 usable cores; v0.2.6 measured **+1.43x on a 4-core
+#: host**, but on a **2-vCPU** host (e.g. a small HuggingFace Space) the 2
+#: clones + GIL contention on the Python-side observation build make it a net
+#: *loss* vs. serial. So we stay serial below the threshold.
+_DEFAULT_MIN_PARALLEL_WORKERS = 4
+
+#: Environment override for operators who know their hardware.
+_MIN_PARALLEL_WORKERS_ENV = "EXPERT_OP4GRID_MIN_PARALLEL_REASSESS_WORKERS"
+
+
+def _min_parallel_workers() -> int:
+    """Minimum worker count to enable parallel reassessment (env-tunable).
+
+    Floored at 2 (a single worker is strictly worse than serial: it pays the
+    clone overhead for zero concurrency). Set the env var high (e.g. 99) to
+    force serial on any host, or to 2 to restore the pre-0.2.7.post1 aggressive
+    behaviour.
+    """
+    raw = os.environ.get(_MIN_PARALLEL_WORKERS_ENV)
+    if raw is None:
+        return _DEFAULT_MIN_PARALLEL_WORKERS
+    try:
+        return max(2, int(raw))
+    except ValueError:
+        return _DEFAULT_MIN_PARALLEL_WORKERS
+
+
+def _should_parallelize_reassessment(is_pypowsybl: bool, workers: int,
+                                     n_actions: int) -> bool:
+    """Whether to run the reassessment in parallel.
+
+    Parallel only pays off above a core threshold — below it the per-worker
+    network-clone overhead makes it slower than the serial path (see
+    :data:`_DEFAULT_MIN_PARALLEL_WORKERS`).
+    """
+    return bool(is_pypowsybl) and n_actions >= 2 and workers >= _min_parallel_workers()
+
 
 def _reassessment_worker_count(n_actions: int) -> Tuple[int, int]:
     """Return ``(cores_available, workers)`` for the reassessment pool.
@@ -265,7 +306,7 @@ def _build_detail(action_id, action, obs_simu_action, info_action):
         sim_out: Dict[str, tuple] = {}
         used_workers = 1
 
-        if is_pypowsybl and workers >= 2 and n_actions >= 2:
+        if _should_parallelize_reassessment(is_pypowsybl, workers, n_actions):
             try:
                 main_nm = obs_simu_defaut._network_manager
                 # Capture the N-1 baseline (contingency + maintenance) variant so
diff --git a/pyproject.toml b/pyproject.toml
index c766ee6f..58cdd80e 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "expert_op4grid_recommender"
-version = "0.2.7"
+version = "0.2.7.post1"
 authors = [
     { name="RTE", email="rte@rte-france.com" },
 ]
diff --git a/tests/test_reassessment_parallel_gate.py b/tests/test_reassessment_parallel_gate.py
new file mode 100644
index 00000000..c7a81fe0
--- /dev/null
+++ b/tests/test_reassessment_parallel_gate.py
@@ -0,0 +1,88 @@
+# Copyright (c) 2025-2026, RTE (https://www.rte-france.com)
+# See AUTHORS.txt
+# This Source Code Form is subject to the terms of the Mozilla Public License, version 2.0.
+# If a copy of the MPL was not distributed with this file,
+# you can obtain one at http://mozilla.org/MPL/2.0/.
+# SPDX-License-Identifier: MPL-2.0
+"""Tests for the reassessment parallel-vs-serial gate (0.2.7.post1).
+
+The parallel per-action reassessment clones the whole network per worker, a
+fixed overhead only amortized above ~3-4 cores. Below the threshold the serial
+path is faster (measured: ~42 s vs ~9 s for the same 15-action pan-European
+case on a 2-vCPU Space vs a many-core Mac). These tests pin the gate so a
+2-vCPU host stays serial while ≥4-core hosts keep the speed-up, plus the env
+override.
+"""
+from __future__ import annotations
+
+import pytest
+
+from expert_op4grid_recommender.utils.reassessment import (
+    _DEFAULT_MIN_PARALLEL_WORKERS,
+    _MIN_PARALLEL_WORKERS_ENV,
+    _min_parallel_workers,
+    _should_parallelize_reassessment,
+)
+
+
+class TestMinParallelWorkers:
+    def test_default_is_four(self, monkeypatch):
+        monkeypatch.delenv(_MIN_PARALLEL_WORKERS_ENV, raising=False)
+        assert _min_parallel_workers() == _DEFAULT_MIN_PARALLEL_WORKERS == 4
+
+    def test_env_override(self, monkeypatch):
+        monkeypatch.setenv(_MIN_PARALLEL_WORKERS_ENV, "2")
+        assert _min_parallel_workers() == 2
+        monkeypatch.setenv(_MIN_PARALLEL_WORKERS_ENV, "99")
+        assert _min_parallel_workers() == 99
+
+    def test_env_floored_at_two(self, monkeypatch):
+        # A single worker is strictly worse than serial → never below 2.
+        monkeypatch.setenv(_MIN_PARALLEL_WORKERS_ENV, "1")
+        assert _min_parallel_workers() == 2
+        monkeypatch.setenv(_MIN_PARALLEL_WORKERS_ENV, "0")
+        assert _min_parallel_workers() == 2
+
+    def test_invalid_env_falls_back_to_default(self, monkeypatch):
+        monkeypatch.setenv(_MIN_PARALLEL_WORKERS_ENV, "not-a-number")
+        assert _min_parallel_workers() == _DEFAULT_MIN_PARALLEL_WORKERS
+
+
+class TestShouldParallelize:
+    def test_two_cores_stays_serial_by_default(self, monkeypatch):
+        # HuggingFace 2-vCPU Space: workers=2 → serial (the fix).
+        monkeypatch.delenv(_MIN_PARALLEL_WORKERS_ENV, raising=False)
+        assert _should_parallelize_reassessment(True, workers=2, n_actions=15) is False
+
+    def test_four_cores_parallelizes(self, monkeypatch):
+        # The v0.2.6-measured positive case (+1.43x on a 4-core host).
+        monkeypatch.delenv(_MIN_PARALLEL_WORKERS_ENV, raising=False)
+        assert _should_parallelize_reassessment(True, workers=4, n_actions=15) is True
+
+    def test_many_cores_parallelizes(self, monkeypatch):
+        monkeypatch.delenv(_MIN_PARALLEL_WORKERS_ENV, raising=False)
+        assert _should_parallelize_reassessment(True, workers=9, n_actions=15) is True
+
+    def test_grid2op_never_parallelizes(self, monkeypatch):
+        monkeypatch.delenv(_MIN_PARALLEL_WORKERS_ENV, raising=False)
+        assert _should_parallelize_reassessment(False, workers=16, n_actions=15) is False
+
+    def test_single_action_stays_serial(self, monkeypatch):
+        monkeypatch.delenv(_MIN_PARALLEL_WORKERS_ENV, raising=False)
+        assert _should_parallelize_reassessment(True, workers=8, n_actions=1) is False
+
+    def test_env_can_force_serial_on_any_host(self, monkeypatch):
+        monkeypatch.setenv(_MIN_PARALLEL_WORKERS_ENV, "99")
+        assert _should_parallelize_reassessment(True, workers=16, n_actions=64) is False
+
+    def test_env_can_restore_aggressive_behaviour(self, monkeypatch):
+        # Set to 2 → the pre-0.2.7.post1 behaviour (parallel from 2 workers).
+        monkeypatch.setenv(_MIN_PARALLEL_WORKERS_ENV, "2")
+        assert _should_parallelize_reassessment(True, workers=2, n_actions=15) is True
+
+
+@pytest.mark.parametrize("workers,expected", [(1, False), (2, False), (3, False),
+                                              (4, True), (8, True)])
+def test_default_threshold_boundary(monkeypatch, workers, expected):
+    monkeypatch.delenv(_MIN_PARALLEL_WORKERS_ENV, raising=False)
+    assert _should_parallelize_reassessment(True, workers=workers, n_actions=15) is expected