diff --git a/scripts/run_tests.sh b/scripts/run_tests.sh
index a1770f934..1a417826f 100644
--- a/scripts/run_tests.sh
+++ b/scripts/run_tests.sh
@@ -13,6 +13,9 @@ SCRIPT_DIR=$(cd "$(dirname "$0")" && pwd)
 REPO_ROOT="$(cd "${SCRIPT_DIR}/.." && pwd)"
 cd "${REPO_ROOT}"
 
+# Broad test runs must be deterministic; dedicated autotune tests opt in per test.
+export FLYDSL_AUTOTUNE=0
+
 # Auto-select GPU with the most free VRAM (skip if HIP_VISIBLE_DEVICES is already set).
 if [[ -z "${HIP_VISIBLE_DEVICES:-}" ]] && command -v python3 &>/dev/null; then
     _best_gpu=$(python3 -c "
diff --git a/tests/unit/test_autotune.py b/tests/unit/test_autotune.py
index 8e97bf605..8fd90d514 100644
--- a/tests/unit/test_autotune.py
+++ b/tests/unit/test_autotune.py
@@ -386,7 +386,8 @@ def default(a, out):
 
 # ── two-track default/search ─────────────────────────────────────────────
 def test_cache_hit_precedes_default_and_search(monkeypatch):
-    monkeypatch.delenv("FLYDSL_AUTOTUNE", raising=False)
+    # The broad test runner uses the explicit off value; search stays opt-in.
+    monkeypatch.setenv("FLYDSL_AUTOTUNE", "0")
     default_calls = 0
 
     def fn(a, out, BLOCK):