diff --git a/src/hyperloom/inference_optimizer/tests/test_integrate_payload_defaults.py b/src/hyperloom/inference_optimizer/tests/test_integrate_payload_defaults.py
index 00f1bf8a78..321b96b2c5 100644
--- a/src/hyperloom/inference_optimizer/tests/test_integrate_payload_defaults.py
+++ b/src/hyperloom/inference_optimizer/tests/test_integrate_payload_defaults.py
@@ -54,7 +54,9 @@ def test_all_three_defaults_fired(self, session_dir):
             session_dir=session_dir,
         )
 
-        assert out["base_tput"] == 800.0
+        # base_tput prefers current_best.tput (900) over the raw baseline (800):
+        # the candidate must beat the recipe it stacks onto, not just baseline.
+        assert out["base_tput"] == 900.0
         assert out["config_path"] == "/tmp/base.yaml"
         assert out["extra_server_args"] == "--page-size 16"
         assert out["kernel_id"] == "k_abc"
@@ -137,6 +139,40 @@ def test_zero_state_does_not_overwrite_explicit_payload(self, session_dir):
 
         assert out["base_tput"] == 750.0
 
+    def test_base_tput_prefers_current_best_over_baseline(self, session_dir):
+        """A candidate must be judged against the CURRENT BEST recipe it stacks
+        onto (current_best.tput), not the raw baseline.
+
+        Otherwise a kernel/fusion that beats baseline but regresses vs the
+        established best (e.g. a warm-replay recipe) is wrongly KEEP'd instead of
+        REVERT'd (observed: forge_fusion adopted at negative gain vs current_best
+        while still positive vs baseline, dragging the final recipe down).
+        """
+        _seed_state(
+            session_dir,
+            baseline_tput=800.0,
+            current_best_args="--page-size 16",  # _seed_state sets current_best.tput=900
+        )
+
+        out = krh._fill_integrate_defaults_from_state(
+            {"kernel_id": "k_abc"},
+            session_dir=session_dir,
+        )
+
+        assert out["base_tput"] == 900.0  # current_best, NOT baseline 800
+
+    def test_base_tput_falls_back_to_baseline_without_current_best(self, session_dir):
+        # No current_best recorded yet (early kernel phase) -> baseline is the
+        # only reference available.
+        _seed_state(session_dir, baseline_tput=800.0)
+
+        out = krh._fill_integrate_defaults_from_state(
+            {"kernel_id": "k_abc"},
+            session_dir=session_dir,
+        )
+
+        assert out["base_tput"] == 800.0
+
 
 class TestIntegrateHandlerHonoursStateDefault:
     @pytest.mark.asyncio
diff --git a/src/hyperloom/orchestrator/kernel/request_handlers.py b/src/hyperloom/orchestrator/kernel/request_handlers.py
index 05e20b629c..487b03e130 100644
--- a/src/hyperloom/orchestrator/kernel/request_handlers.py
+++ b/src/hyperloom/orchestrator/kernel/request_handlers.py
@@ -1264,8 +1264,19 @@ def _fill_integrate_defaults_from_state(
             str(pending_record.get("identity_route") or ""),
         )
 
+    current_best = getattr(state, "current_best", None) or {}
+
     if float(resolved.get("base_tput", 0.0) or 0.0) <= 0:
-        bt = float(getattr(state, "baseline_tput", 0.0) or 0.0)
+        # Judge the candidate against the CURRENT BEST recipe it stacks onto,
+        # not the raw baseline. ``extra_server_args`` below is filled from
+        # current_best, so the candidate is re-benched on top of that recipe;
+        # comparing the result to the raw baseline lets a kernel/fusion that
+        # beats baseline but REGRESSES vs the established best (e.g. a
+        # warm-replay recipe) get KEEP'd and drag the recipe down. Mirrors
+        # integrate_patch's rebind to current_best. Baseline is the fallback
+        # only before any current_best exists.
+        cb_tput = float(current_best.get("tput") or 0.0) if isinstance(current_best, dict) else 0.0
+        bt = cb_tput if cb_tput > 0 else float(getattr(state, "baseline_tput", 0.0) or 0.0)
         if bt > 0:
             resolved["base_tput"] = bt
 
@@ -1274,7 +1285,6 @@ def _fill_integrate_defaults_from_state(
         if cfg:
             resolved["config_path"] = cfg
 
-    current_best = getattr(state, "current_best", None) or {}
     if not resolved.get("extra_server_args") and isinstance(current_best, dict):
         cb_args = current_best.get("extra_server_args") or ""
         if cb_args: