code/serving/patches/jevbench-image-adapter.patch

diff --git a/jevbench/adapters/__init__.py b/jevbench/adapters/__init__.py
index c71d476..6aacaa8 100644
--- a/jevbench/adapters/__init__.py
+++ b/jevbench/adapters/__init__.py
@@ -17,6 +17,7 @@ Raw provider responses are preserved by the runner, not here.
 
 from .base import DecisionResult  # noqa: F401
 from .typesafe import TypeSafeAdapter  # noqa: F401
+from .openjev_image import OpenJevImageAdapter  # noqa: F401
 from .openai_compat import OpenAICompatAdapter  # noqa: F401
 from .systemone_list import SystemOneListAdapter  # noqa: F401
 from .gradio_space import GradioSpaceAdapter  # noqa: F401
diff --git a/jevbench/adapters/openjev_image.py b/jevbench/adapters/openjev_image.py
new file mode 100644
index 0000000..7aab898
--- /dev/null
+++ b/jevbench/adapters/openjev_image.py
@@ -0,0 +1,30 @@
+"""OpenJev image Decisions transport; normalization is disclosed, never fitted.
+
+Task state: {"text": "An image: <<IMG>>", "image_data": "<base64>"}.
+Keeping the image bytes in the task binds them into the existing dataset hash.
+Uses the unchanged JevBench scoring/ledger/raw-response path.
+"""
+from .typesafe import TypeSafeAdapter
+
+
+class OpenJevImageAdapter(TypeSafeAdapter):
+    name = "openjev_image"
+
+    def build_request(self, task):
+        state = task.state
+        if not isinstance(state, dict) or not isinstance(state.get("image_data"), str):
+            raise ValueError("openjev_image requires state.image_data as base64 or a data URI")
+        if not isinstance(state.get("text", ""), str):
+            raise ValueError("state.text must be text")
+        body = super().build_request(task)
+        body["state"] = state.get("text", "An image: <<IMG>>")
+        body["image_data"] = state["image_data"]
+        return body
+
+    def run(self, task):
+        result = super().run(task)
+        result.probs_source = "normalized_entailment_v1"
+        if result.ok and result.raw.get("probability_method") != "normalized_entailment_v1":
+            result.ok = False
+            result.error = "Endpoint did not declare normalized_entailment_v1"
+        return result
diff --git a/jevbench/cli.py b/jevbench/cli.py
index da38bea..fb79845 100644
--- a/jevbench/cli.py
+++ b/jevbench/cli.py
@@ -23,7 +23,7 @@ import time
 from .adapters import (GradioSpaceAdapter, LocalOpenJevAdapter, NeedleLocalAdapter,
                        OpenAICompatAdapter, SystemOneListAdapter,
                        RemoteInprocAdapter, SemIfDirectAdapter, SgSystemOneAdapter, So1DeciderAdapter,
-                       TypeSafeAdapter, DjevAdapter,
+                       TypeSafeAdapter, OpenJevImageAdapter, DjevAdapter,
                        LayaLocalAdapter, Gliner2LocalAdapter, VerdictLocalAdapter, PawLocalAdapter,
                        ClassifierDevAdapter, CertoLocalAdapter, QwenFlashLinearAdapter)
 from .budget import Ledger
@@ -55,7 +55,7 @@ def cmd_run(args) -> int:
     if args.limit:
         tasks = tasks[: args.limit]
 
-    kinds = {"typesafe": TypeSafeAdapter, "systemone_list": SystemOneListAdapter,
+    kinds = {"typesafe": TypeSafeAdapter, "openjev_image": OpenJevImageAdapter, "systemone_list": SystemOneListAdapter,
              "gradio_space": GradioSpaceAdapter, "local_openjev": LocalOpenJevAdapter,
              "openai_compat": OpenAICompatAdapter,
              "needle_local": NeedleLocalAdapter,
@@ -161,7 +161,7 @@ def main(argv=None) -> int:
     p_run = sub.add_parser("run", help="run the benchmark over tasks")
     p_run.add_argument("--tasks", required=True)
     p_run.add_argument("--adapter", required=True,
-                       choices=["typesafe", "systemone_list", "gradio_space",
+                       choices=["typesafe", "openjev_image", "systemone_list", "gradio_space",
                                 "local_openjev", "openai_compat", "needle_local",
                                 "semif_direct", "so1_decider", "remote_inproc",
                                 "sg_system_one", "djev", "laya_local", "gliner2_local",