diff --git a/runs/frontier-multicase-sufficiency-v0/best_effort/frontier_moe_ep_feature_contract.patch b/runs/frontier-multicase-sufficiency-v0/best_effort/frontier_moe_ep_feature_contract.patch
new file mode 100644
index 0000000..70a1c7a
--- /dev/null
+++ b/runs/frontier-multicase-sufficiency-v0/best_effort/frontier_moe_ep_feature_contract.patch
@@ -0,0 +1,75 @@
+diff --git a/frontier/execution_time_predictor/sklearn_moe_execution_time_predictor.py b/frontier/execution_time_predictor/sklearn_moe_execution_time_predictor.py
+--- a/frontier/execution_time_predictor/sklearn_moe_execution_time_predictor.py
++++ b/frontier/execution_time_predictor/sklearn_moe_execution_time_predictor.py
+@@ -424 +424,4 @@
+-            features = self._build_moe_load_imbalance_features(allocation)
++            features = self._build_moe_load_imbalance_features(
++                allocation,
++                num_tokens=total_routed_tokens // self._router_topk,
++            )
+@@ -1365,0 +1369 @@
++        num_tokens: Optional[int] = None,
+@@ -1384,3 +1388,6 @@
+-        approx_num_tokens = max(
+-            1, int(round(total_routed_tokens / float(self._router_topk)))
+-        )
++        if num_tokens is None:
++            num_tokens = max(
++                1, int(round(total_routed_tokens / float(self._router_topk)))
++            )
++        elif num_tokens <= 0:
++            raise ValueError(f"num_tokens must be positive, got {num_tokens}")
+@@ -1389 +1396 @@
+-            num_tokens=approx_num_tokens,
++            num_tokens=int(num_tokens),
+@@ -1475,0 +1483,3 @@
++                profile_num_tokens = int(
++                    self._get_effective_moe_total_tokens(batch)
++                )
+@@ -1478,0 +1489 @@
++                    profile_num_tokens,
+@@ -1486 +1497,2 @@
+-                        per_expert_tokens
++                        per_expert_tokens,
++                        num_tokens=profile_num_tokens,
+@@ -1778,4 +1790,4 @@
+-            # The grouped_gemm predictor expects pre-routing num_tokens metadata, while runtime
+-            # allocation is post-routing (already expanded by router_topk). Recover the
+-            # approximate pre-routing token count to keep feature semantics aligned with
+-            # the profiling dataset contract.
++            # The grouped_gemm predictor keeps global pre-routing num_tokens and
++            # lane-local routed assignments as independent features. Inferring the
++            # former from the latter is valid for EP=1 but undercounts by EP size
++            # for expert-parallel execution.
+@@ -1784,2 +1796,7 @@
+-            approx_num_tokens = max(
+-                1, int(round(total_routed_tokens / float(self._router_topk)))
++            profile_num_tokens = (
++                int(self._get_effective_moe_total_tokens(batch))
++                if batch is not None
++                else max(
++                    1,
++                    int(round(total_routed_tokens / float(self._router_topk))),
++                )
+@@ -1804 +1821 @@
+-                approx_num_tokens,
++                profile_num_tokens,
+@@ -1811 +1828,4 @@
+-                features = self._build_moe_load_imbalance_features(per_expert_tokens)
++                features = self._build_moe_load_imbalance_features(
++                    per_expert_tokens,
++                    num_tokens=profile_num_tokens,
++                )
+@@ -1851,2 +1871,7 @@
+-            approx_num_tokens = max(
+-                1, int(round(total_routed_tokens / float(self._router_topk)))
++            profile_num_tokens = (
++                int(self._get_effective_moe_total_tokens(batch))
++                if batch is not None
++                else max(
++                    1,
++                    int(round(total_routed_tokens / float(self._router_topk))),
++                )
+@@ -1854 +1879 @@
+-            rounded_tokens = self._round_to_valid_key(approx_num_tokens)
++            rounded_tokens = self._round_to_valid_key(profile_num_tokens)