diff --git a/test/evals/evaluators.py b/test/evals/evaluators.py
index de301aaed3e..9501c9728e2 100644
--- a/test/evals/evaluators.py
+++ b/test/evals/evaluators.py
@@ -1,13 +1,22 @@
 """Custom pydantic-evals evaluators for Galaxy agents."""
 
 from dataclasses import dataclass
-from typing import Any
+from typing import (
+    Any,
+    Generic,
+    TypeVar,
+)
 
 from pydantic_evals.evaluators import (
     Evaluator,
     EvaluatorContext,
 )
 
+# Routing datasets vary in their case-input shape: a plain query string, or a dict for the
+# multi-turn cases. HandoffMatch ignores the input entirely (it only compares output to
+# expected), so it is generic over the input type and adapts to whichever dataset it scores.
+HandoffInputsT = TypeVar("HandoffInputsT")
+
 
 def _output_text(output: Any) -> str:
     """Pull a string out of a task output that may be a str or {"content": str, ...} dict."""
@@ -17,13 +26,10 @@ def _output_text(output: Any) -> str:
 
 
 @dataclass
-class HandoffMatch(Evaluator[Any, str, dict]):
-    """Score 1.0 if router's chosen agent_type matches expected, 0.0 otherwise.
+class HandoffMatch(Evaluator[HandoffInputsT, str, dict], Generic[HandoffInputsT]):
+    """Score 1.0 if router's chosen agent_type matches expected, 0.0 otherwise."""
 
-    Input type is ``Any`` so this works with both string-query datasets (routing) and
-    dict-input multi-turn datasets (routing_depth, routing_clarification_followup)."""
-
-    def evaluate(self, ctx: EvaluatorContext[Any, str, dict]) -> float:
+    def evaluate(self, ctx: EvaluatorContext[HandoffInputsT, str, dict]) -> float:
         return 1.0 if ctx.output == ctx.expected_output else 0.0
 
 
diff --git a/test/evals/specs.py b/test/evals/specs.py
index 3faaec7111e..1c8d269aade 100644
--- a/test/evals/specs.py
+++ b/test/evals/specs.py
@@ -10,7 +10,9 @@ from collections.abc import (
 from dataclasses import dataclass
 from typing import (
     Any,
+    Generic,
     Optional,
+    TypeVar,
 )
 
 from pydantic_ai.models import Model
@@ -47,17 +49,18 @@ from .tasks import (
     make_tool_recommendation_task,
 )
 
+# A case input is either a plain query string or, for the multi-turn datasets
+# (routing_depth, routing_clarification_followup), a dict. Parameterizing BuiltDataset on it
+# keeps the dataset and its task linked: a dict-input dataset must pair with a dict-input task.
+CaseInputsT = TypeVar("CaseInputsT")
+
 
 @dataclass
-class BuiltDataset:
-    """A dataset configured with evaluators and a task ready to evaluate.
+class BuiltDataset(Generic[CaseInputsT]):
+    """A dataset configured with evaluators and a task ready to evaluate."""
 
-    Input type is ``Any``: most datasets take a string query, but the multi-turn ones
-    (routing_depth, routing_clarification_followup) take a dict case input.
-    """
-
-    dataset: Dataset[Any, Any, dict[str, Any]]
-    task: Callable[..., Awaitable[Any]]
+    dataset: Dataset[CaseInputsT, Any, dict[str, Any]]
+    task: Callable[[CaseInputsT], Awaitable[Any]]
     primary_score: str  # name of the headline scorer for the summary table