diff --git a/CHANGELOG.md b/CHANGELOG.md
index 21b9dd7..e17961c 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -4,6 +4,11 @@ The format follows [Keep a Changelog](https://keepachangelog.com/en/1.1.0/); ver
 
 ## Unreleased
 
+### Fixed
+
+- Rail evaluations charge Jev input tokens only for the Jev backend. Local Laya evaluations keep
+  `jev_input_tokens` and `cost_usd` at zero, matching the tool and browser fronts.
+
 ### Added
 
 - `docs/benchmarks.md`: the Google Flights driver comparison rerun on 2026-09-23 from Poland, every arm three times on
diff --git a/s1a/rails.py b/s1a/rails.py
index b8941da..2935f18 100644
--- a/s1a/rails.py
+++ b/s1a/rails.py
@@ -120,7 +120,8 @@ async def evaluate(
     tp = sum(a and label for a, label in zip(acted, labels))
     fp = sum(a and not label for a, label in zip(acted, labels))
     fn = sum(label and not a for a, label in zip(acted, labels))
-    jev_input_tokens = sum(verdict.input_tokens for verdict in verdicts)
+    # Local models have no Jev API charges, as on the tool and browser fronts.
+    jev_input_tokens = sum(verdict.input_tokens for verdict in verdicts) if decision_model.name == "jev" else 0
     summary = {
         "rail": spec.name,
         "records": len(records),
diff --git a/tests/test_rails.py b/tests/test_rails.py
index 2fa1fff..a5f10dc 100644
--- a/tests/test_rails.py
+++ b/tests/test_rails.py
@@ -18,7 +18,7 @@
 
 from s1a import rails
 from s1a.agents.injection_guard import SPEC as GUARD
-from s1a.decision_models import JevModel, ScriptedModel, ScriptedTransport, Usage
+from s1a.decision_models import JevModel, LayaModel, ScriptedModel, ScriptedTransport, Usage
 from s1a.spec import RailSpec, Thresholds
 
 INJECTED = (
@@ -204,7 +204,12 @@ async def test_precision_and_recall_of_the_act_band_against_the_labels(self) ->
             labelled = Path(tmp) / "set.jsonl"
             labelled.write_text("".join(json.dumps(r) + "\n" for r in records), encoding="utf-8")
             summary = await rails.evaluate(
-                GUARD, labelled, decision_model=_noul([0.1, 0.9, 0.95, 0.2, 0.5]), results_dir=Path(tmp) / "results"
+                GUARD,
+                labelled,
+                decision_model=JevModel(
+                    ScriptedTransport(noul=[0.1, 0.9, 0.95, 0.2, 0.5], usage={"input_tokens": 300}, latency_ms=7)
+                ),
+                results_dir=Path(tmp) / "results",
             )
             job_dir = Path(summary["job_dir"])
             verdicts = [json.loads(line) for line in (job_dir / "verdicts.jsonl").read_text().splitlines()]
@@ -217,7 +222,29 @@ async def test_precision_and_recall_of_the_act_band_against_the_labels(self) ->
         self.assertEqual([v["band"] for v in verdicts], ["allow", "act", "act", "allow", "uncertain"])
         self.assertEqual(written["rail"], "injection_guard")
         self.assertEqual(job_dir.parent, Path(tmp) / "results" / "injection_guard")
-        self.assertTrue(job_dir.name.endswith("__scripted"))  # the model's name, jev or laya on a real run
+        self.assertTrue(job_dir.name.endswith("__jev"))
+
+    async def test_laya_usage_is_not_charged_as_jev_in_the_returned_or_saved_summary(self) -> None:
+        agent = SimpleNamespace(
+            cfg={"max_len": 512},
+            system_one=lambda state, questions: {
+                "answers": {"check": {"noul": 0.9}},
+                "usage": {"input_tokens": 300},
+            },
+        )
+        decision_model = LayaModel(agent, model="laya-test")
+        verdict = await rails.ask(GUARD, {"text": INJECTED}, decision_model)
+        self.assertEqual(verdict.input_tokens, 300)
+        with tempfile.TemporaryDirectory() as tmp:
+            labelled = Path(tmp) / "set.jsonl"
+            labelled.write_text(json.dumps({"state": {"text": INJECTED}, "label": True}) + "\n", encoding="utf-8")
+            summary = await rails.evaluate(GUARD, labelled, decision_model=decision_model, results_dir=Path(tmp))
+            job_dir = Path(summary["job_dir"])
+            written = json.loads((job_dir / "summary.json").read_text(encoding="utf-8"))
+        for result in (summary, written):
+            self.assertEqual((result["jev_input_tokens"], result["cost_usd"]), (0, 0.0))
+            self.assertEqual((result["records"], result["accuracy"]), (1, 1.0))
+        self.assertTrue(job_dir.name.endswith("__laya"))
 
     def test_a_record_without_a_boolean_label_is_rejected(self) -> None:
         with tempfile.TemporaryDirectory() as tmp: