diff --git a/hud/cli/eval.py b/hud/cli/eval.py
index 8b3f4ccd1..1530720be 100644
--- a/hud/cli/eval.py
+++ b/hud/cli/eval.py
@@ -301,6 +301,15 @@ def _parse_agent_type(cls, v: Any) -> AgentType | None:
             try:
                 return AgentType(v)
             except ValueError:
+                if v == "integration_test":
+                    raise ValueError(
+                        "integration_test is the platform's authoring agent: "
+                        "it pre-stages the golden solution (Task.validation) "
+                        "and runs the graders expecting Reward 1.0. Run it "
+                        "against a deployed environment on the platform — "
+                        "`hud eval <env-name> integration_test --task-ids "
+                        "<slug> -y` — it is not available with --runtime local."
+                    ) from None
                 valid = [e.value for e in AgentType]
                 raise ValueError(
                     f"Invalid agent: {v}. Must be one of: {', '.join(valid)}"
diff --git a/hud/cli/tests/test_eval_config.py b/hud/cli/tests/test_eval_config.py
index fad19d88f..62a7556ee 100644
--- a/hud/cli/tests/test_eval_config.py
+++ b/hud/cli/tests/test_eval_config.py
@@ -25,6 +25,11 @@ def test_is_bedrock_arn() -> None:
     assert _is_bedrock_arn(None) is False
 
 
+def test_parse_agent_type_points_integration_test_to_the_platform() -> None:
+    with pytest.raises(ValueError, match="integration_test is the platform"):
+        EvalConfig(agent_type="integration_test")
+
+
 def test_parse_agent_type_accepts_known_value() -> None:
     cfg = EvalConfig(agent_type="openai")
     assert cfg.agent_type is not None
diff --git a/hud/integrations/harbor/adapt.py b/hud/integrations/harbor/adapt.py
index 81b86bdf2..eae0ec3c7 100644
--- a/hud/integrations/harbor/adapt.py
+++ b/hud/integrations/harbor/adapt.py
@@ -189,6 +189,13 @@ class TaskConfig(BaseModel):
     steps: list[dict[str, Any]] | None = None
 
 
+# The Harbor task schema this build adapts. export() stamps the same value
+# into generated task.toml files so round-tripped tasks validate; a task that
+# declares any other schema_version was authored against a different adapter
+# and must fail loudly rather than adapt with silently wrong semantics.
+HARBOR_SCHEMA_VERSION = "1.0"
+
+
 @dataclass(frozen=True, slots=True)
 class HarborTask:
     path: Path
@@ -238,6 +245,12 @@ def adapt(
             raise ValueError(
                 f"{task_dir.name}/task.toml is not a valid Harbor task: {error}"
             ) from error
+        if config.schema_version is not None and config.schema_version != HARBOR_SCHEMA_VERSION:
+            raise ValueError(
+                f"{task_dir.name}/task.toml declares unsupported Harbor schema "
+                f"{config.schema_version!r} — this HUD build adapts schema "
+                f"{HARBOR_SCHEMA_VERSION!r}"
+            )
         unsupported = []
         if config.environment.os != "linux":
             unsupported.append(f"os={config.environment.os!r}")
diff --git a/hud/integrations/harbor/export.py b/hud/integrations/harbor/export.py
index 4eb4bbe7c..b304a6fd6 100644
--- a/hud/integrations/harbor/export.py
+++ b/hud/integrations/harbor/export.py
@@ -12,6 +12,7 @@
 from hud.environment import Environment, load_environment
 from hud.environment.server import TaskRunner
 from hud.eval import Taskset
+from hud.integrations.harbor.adapt import HARBOR_SCHEMA_VERSION
 from hud.utils.naming import normalize_environment_name
 
 ALLOWED_PROTOCOLS = ("ssh", "mcp")
@@ -162,6 +163,7 @@ def ignore_export(dirpath: str, names: list[str]) -> set[str]:
             args_json = json.dumps(task.args)
             (task_dir / "task.toml").write_text(
                 'version = "1.0"\n'
+                f'schema_version = "{HARBOR_SCHEMA_VERSION}"\n'
                 f"name = {json.dumps(slug)}\n"
                 "\n[metadata]\n"
                 f"hud_task = {json.dumps(task.id)}\n"
diff --git a/hud/integrations/harbor/tests/test_contract.py b/hud/integrations/harbor/tests/test_contract.py
index 84fb6a14f..98704d999 100644
--- a/hud/integrations/harbor/tests/test_contract.py
+++ b/hud/integrations/harbor/tests/test_contract.py
@@ -12,6 +12,7 @@
 
 from hud.eval import Task
 from hud.integrations import harbor
+from hud.integrations.harbor.adapt import HARBOR_SCHEMA_VERSION
 
 from .conftest import make_harbor_task, make_multi_step_task
 
@@ -1057,3 +1058,27 @@ def test_authored_runtime_assets_are_valid_source() -> None:
 
 def test_public_surface_is_only_the_two_real_operations() -> None:
     assert harbor.__all__ == ["adapt", "export"]
+
+
+def test_unknown_schema_version_fails_loudly(tmp_path: Path) -> None:
+    task = make_harbor_task(tmp_path, "task-a")
+    (task / "task.toml").write_text('schema_version = "99.9"\n', encoding="utf-8")
+
+    with pytest.raises(ValueError, match="unsupported Harbor schema"):
+        harbor.adapt(tmp_path)
+
+
+def test_supported_schema_version_adapts(tmp_path: Path) -> None:
+    task = make_harbor_task(tmp_path, "task-a")
+    (task / "task.toml").write_text(
+        f'schema_version = "{HARBOR_SCHEMA_VERSION}"\n', encoding="utf-8"
+    )
+
+    taskset = harbor.adapt(tmp_path)
+    assert len(list(taskset)) == 1
+
+
+def test_absent_schema_version_still_adapts(tmp_path: Path) -> None:
+    """Unversioned tasks (the historical export shape) keep adapting."""
+    make_harbor_task(tmp_path, "task-a")
+    assert len(list(harbor.adapt(tmp_path))) == 1