diff --git a/app/desktop/studio_server/test_eval_api.py b/app/desktop/studio_server/test_eval_api.py
index 8ea0cb9a5a..8b20306c88 100644
--- a/app/desktop/studio_server/test_eval_api.py
+++ b/app/desktop/studio_server/test_eval_api.py
@@ -6388,6 +6388,33 @@ def test_exact_match_fail(self, client, mock_task, mock_task_from_id):
         assert response.status_code == 200
         assert response.json()["scores"]["accuracy"] == 0.0
 
+    @pytest.mark.parametrize(
+        "final_message,expected_score",
+        [
+            # Same expression and same expected_value both times. The only
+            # difference is whether `user` is there, so the 0.0 can only come
+            # from the extraction failing -- not from a value that never matched.
+            ('{"user": {"status": "hello"}}', 1.0),
+            ('{"status": "hello"}', 0.0),
+        ],
+        ids=["field_present", "field_missing"],
+    )
+    def test_missing_nested_field_scores_fail(
+        self, client, mock_task, mock_task_from_id, final_message, expected_score
+    ):
+        # Reaching into a value that isn't there is a scored FAIL, not a request error.
+        mock_task_from_id.return_value = mock_task
+        payload = self._payload()
+        payload["properties"]["value_expression"] = (
+            "(final_message | fromjson).user.status"
+        )
+        payload["eval_input"]["final_message"] = final_message
+        response = client.post(self._url(), json=payload)
+        assert response.status_code == 200
+        body = response.json()
+        assert body["scores"]["accuracy"] == expected_score
+        assert body["skipped_reason"] is None
+
     def test_nothing_is_persisted(self, client, mock_task, mock_task_from_id):
         mock_task_from_id.return_value = mock_task
         response = client.post(self._url(), json=self._payload())
diff --git a/libs/core/kiln_ai/adapters/eval/test_v2_eval_helpers.py b/libs/core/kiln_ai/adapters/eval/test_v2_eval_helpers.py
index f366044a5f..bba4ca02cf 100644
--- a/libs/core/kiln_ai/adapters/eval/test_v2_eval_helpers.py
+++ b/libs/core/kiln_ai/adapters/eval/test_v2_eval_helpers.py
@@ -170,6 +170,19 @@ def test_fromjson_valid_json_succeeds(self):
         assert value == "val"
         assert fail_result is None
 
+    def test_empty_trace_scores_fail_not_missing_trace_skip(self):
+        # A trace that exists but is empty is not a missing trace: the run did
+        # produce one, it just has nothing in it. That scores FAIL, so the
+        # missing-trace skip must key off `is None`, not falsiness.
+        inp = _make_input(trace=[])
+        value, fail_result = extract_output_value(
+            "trace[-1].tool_calls[0].function.name", inp, _SAMPLE_SCORES
+        )
+        assert value is None
+        assert fail_result is not None
+        assert fail_result.skipped_reason is None
+        assert fail_result.scores == {"s1": 0.0, "s2": 0.0}
+
 
 # ---------------------------------------------------------------------------
 # stringify_for_match
diff --git a/libs/core/kiln_ai/datamodel/eval.py b/libs/core/kiln_ai/datamodel/eval.py
index 7d0cc71823..03e4ca642c 100644
--- a/libs/core/kiln_ai/datamodel/eval.py
+++ b/libs/core/kiln_ai/datamodel/eval.py
@@ -1028,7 +1028,7 @@ def validate_properties(self) -> Self:
             raise ValueError(f"Invalid eval config type: {self.config_type}")
 
     @model_validator(mode="after")
-    def validate_v2_templates_and_expressions(self) -> Self:
+    def validate_v2_templates_and_expressions(self, info: ValidationInfo) -> Self:
         if self.config_type != EvalConfigType.v2 or not isinstance(
             self.properties, BaseModel
         ):
@@ -1037,6 +1037,7 @@ def validate_v2_templates_and_expressions(self) -> Self:
         from kiln_ai.utils.jinja_engine import (
             compile_expression_or_raise,
             compile_template_or_raise,
+            expression_variables,
         )
 
         props = self.properties
@@ -1069,6 +1070,25 @@ def validate_v2_templates_and_expressions(self) -> Self:
         ):
             if props.value_expression is not None:
                 compile_expression_or_raise(props.value_expression)
+                # Syntax alone isn't enough: a typo'd root variable resolves to
+                # Undefined at eval time, which scores every row a silent 0.0.
+                # Catch it while the author can still see what they typed.
+                #
+                # Authoring-time only. A config written before this check exists
+                # may name a variable we now reject, and refusing to load it would
+                # take the whole eval down rather than the one check that was
+                # already scoring zeros.
+                if not self.loading_from_file(info):
+                    allowed = set(EvalTaskInput.model_fields.keys())
+                    unknown = sorted(
+                        expression_variables(props.value_expression) - allowed
+                    )
+                    if unknown:
+                        raise ValueError(
+                            f"value_expression references unknown variable "
+                            f"'{unknown[0]}'. Available variables: "
+                            f"{', '.join(sorted(allowed))}."
+                        )
 
         return self
 
diff --git a/libs/core/kiln_ai/datamodel/test_eval_model.py b/libs/core/kiln_ai/datamodel/test_eval_model.py
index 02586cc43b..66487927bf 100644
--- a/libs/core/kiln_ai/datamodel/test_eval_model.py
+++ b/libs/core/kiln_ai/datamodel/test_eval_model.py
@@ -2557,6 +2557,16 @@ def _make_v2_eval_config(**kwargs) -> EvalConfig:
     return EvalConfig(name="V2 Test", config_type=EvalConfigType.v2, **kwargs)
 
 
+# One value_expression per EvalTaskInput field, keyed by the root variable it
+# reads. Checked against EvalTaskInput.model_fields so the coverage claim holds.
+_EVAL_INPUT_FIELD_EXPRESSIONS = {
+    "final_message": "final_message",
+    "trace": "trace[-1].content",
+    "task_input": "task_input | upper",
+    "reference_data": "reference_data",
+}
+
+
 class TestV2TemplateValidation:
     def test_valid_prompt_template(self):
         """A prompt_template with a Jinja expression passes validation."""
@@ -2666,6 +2676,69 @@ def test_invalid_value_expression(self):
                 ),
             )
 
+    @pytest.mark.parametrize(
+        "root,expression",
+        list(_EVAL_INPUT_FIELD_EXPRESSIONS.items()),
+    )
+    def test_value_expression_accepts_every_eval_input_field(self, root, expression):
+        # Tied to the model so a new EvalTaskInput field can't leave this test
+        # named "every field" while silently skipping one.
+        assert set(_EVAL_INPUT_FIELD_EXPRESSIONS) == set(EvalTaskInput.model_fields)
+        cfg = _make_v2_eval_config(
+            properties=ExactMatchProperties(
+                expected_value="yes",
+                value_expression=expression,
+            ),
+        )
+        assert cfg.properties.value_expression == expression
+
+    @pytest.mark.parametrize(
+        "expression,unknown",
+        [
+            ("outpt.status", "outpt"),
+            ("messages[-1].content", "messages"),
+            ("final_message ~ typo", "typo"),
+        ],
+    )
+    def test_value_expression_rejects_unknown_variable(self, expression, unknown):
+        """A typo'd root fails silently at runtime, so it has to fail loudly here.
+
+        Depending on what the expression does with it, the typo either resolves
+        to Undefined and scores every row 0.0, or -- as with `~`, which
+        stringifies Undefined to '' -- produces a plausible-looking wrong value
+        ('hi' for `final_message ~ typo`) that no one notices.
+        """
+        with pytest.raises(
+            ValidationError, match=f"unknown variable '{unknown}'"
+        ) as exc:
+            _make_v2_eval_config(
+                properties=ExactMatchProperties(
+                    expected_value="yes",
+                    value_expression=expression,
+                ),
+            )
+        assert "final_message, reference_data, task_input, trace" in str(exc.value)
+
+    def test_value_expression_unknown_variable_still_loads_from_file(self):
+        """Already-saved configs keep loading; the check gates writes, not reads."""
+        cfg = EvalConfig.model_validate(
+            {
+                "v": 1,
+                "id": "123",
+                "name": "Saved before the check existed",
+                "config_type": "v2",
+                "model_type": "eval_config",
+                "properties": {
+                    "type": "exact_match",
+                    "expected_value": "yes",
+                    "value_expression": "outpt.status",
+                },
+            },
+            context={"loading_from_file": True},
+        )
+        assert isinstance(cfg.properties, ExactMatchProperties)
+        assert cfg.properties.value_expression == "outpt.status"
+
     def test_none_value_expression_skipped(self):
         """value_expression=None (default) should not be validated."""
         cfg = _make_v2_eval_config(
diff --git a/libs/core/kiln_ai/utils/jinja_engine.py b/libs/core/kiln_ai/utils/jinja_engine.py
index ebea2cab70..5ce5bec79f 100644
--- a/libs/core/kiln_ai/utils/jinja_engine.py
+++ b/libs/core/kiln_ai/utils/jinja_engine.py
@@ -10,6 +10,7 @@
 Public API:
   - compile_template_or_raise(template) -> None
   - compile_expression_or_raise(expression) -> None
+  - expression_variables(expression) -> set[str]
   - render_input_transform(transform, task_input) -> str
   - extract(expression, data) -> Any
 """
@@ -20,7 +21,7 @@
 import types
 from typing import TYPE_CHECKING, Any
 
-from jinja2 import StrictUndefined, TemplateSyntaxError, Undefined
+from jinja2 import StrictUndefined, TemplateSyntaxError, Undefined, UndefinedError
 from jinja2.sandbox import SandboxedEnvironment
 
 if TYPE_CHECKING:
@@ -28,10 +29,13 @@
 
 
 class JinjaExtractionError(ValueError):
-    """Raised when a Jinja2 filter encounters invalid data during extraction.
+    """Raised when a Jinja2 expression encounters missing or invalid data during extraction.
 
-    For example, the ``fromjson`` filter raises this when the input string is
-    not valid JSON.
+    Covers both filters that reject their input (``fromjson`` on a non-JSON
+    string) and expressions that operate on a value that isn't there (reaching
+    into a missing nested field, indexing past the end of a list). Callers treat
+    it as "this input didn't have what the expression asked for" rather than an
+    unexpected crash.
     """
 
 
@@ -85,6 +89,24 @@ def compile_expression_or_raise(expression: str) -> None:
         ) from e
 
 
+def expression_variables(expression: str) -> set[str]:
+    """Return the namespace variables a Jinja2 expression reads.
+
+    An expression is not a template -- ``final_message.strip()`` parsed as a
+    template is just literal text with no variables -- so it is wrapped in an
+    output block before the AST walk. Raises ValueError on syntax error.
+    """
+    from jinja2 import meta
+
+    try:
+        ast = _expression_env.parse("{{ " + expression + " }}")
+    except TemplateSyntaxError as e:
+        raise ValueError(
+            f"Invalid Jinja2 expression: {e.message} (line {e.lineno})"
+        ) from e
+    return meta.find_undeclared_variables(ast)
+
+
 def render_input_transform(
     transform: InputTransform,
     task_input: Any,
@@ -108,6 +130,7 @@ def extract(expression: str, data: dict) -> Any:
     - Missing keys return Undefined (not None, not a raise).
     - Explicit null values return None.
     - Generators are auto-materialized to lists.
+    - Operating on a missing value raises JinjaExtractionError, never UndefinedError.
     """
     try:
         compiled = _expression_env.compile_expression(
@@ -117,9 +140,19 @@ def extract(expression: str, data: dict) -> Any:
         raise ValueError(
             f"Invalid Jinja2 expression: {e.message} (line {e.lineno})"
         ) from e
-    result = compiled(**data)
-    if isinstance(result, types.GeneratorType):
-        result = list(result)
+    try:
+        result = compiled(**data)
+        # Materializing inside the try is load-bearing: map/selectattr/groupby
+        # return lazy generators, so their per-item lookups don't run until
+        # list() does. A missing field there raises here, not above.
+        if isinstance(result, types.GeneratorType):
+            result = list(result)
+    except UndefinedError as e:
+        # A single missing lookup yields Undefined, but any further operation on
+        # it (attribute access, indexing, a filter) raises. That's still just
+        # missing data, so surface it as an extraction error callers handle --
+        # keeping Jinja's message, which names the field that wasn't there.
+        raise JinjaExtractionError(str(e)) from e
     return result
 
 
diff --git a/libs/core/kiln_ai/utils/test_jinja_engine.py b/libs/core/kiln_ai/utils/test_jinja_engine.py
index 20d3d37ad7..3e314a2290 100644
--- a/libs/core/kiln_ai/utils/test_jinja_engine.py
+++ b/libs/core/kiln_ai/utils/test_jinja_engine.py
@@ -7,6 +7,7 @@
     JinjaExtractionError,
     compile_expression_or_raise,
     compile_template_or_raise,
+    expression_variables,
     extract,
     render_input_transform,
 )
@@ -173,6 +174,86 @@ def test_malformed_expression_raises_value_error(self):
             extract("{{ invalid", {})
 
 
+class TestExpressionVariables:
+    @pytest.mark.parametrize(
+        "expression,expected",
+        [
+            # A bare expression has no {{ }} of its own, so it only reports
+            # variables once the function wraps it.
+            ("final_message", {"final_message"}),
+            ("final_message.strip()", {"final_message"}),
+            ("(final_message | fromjson).user.status", {"final_message"}),
+            ("trace | map(attribute='x') | list", {"trace"}),
+            ("outpt.status", {"outpt"}),
+            # Delimiters inside a string literal are lexed as text, not as the
+            # end of the wrapper block.
+            ('final_message ~ "}}"', {"final_message"}),
+            ("'literal'", set()),
+        ],
+    )
+    def test_reports_referenced_variables(self, expression, expected):
+        assert expression_variables(expression) == expected
+
+    def test_syntax_error_raises_value_error(self):
+        with pytest.raises(ValueError, match="Invalid Jinja2 expression"):
+            expression_variables("final_message[")
+
+
+class TestExtractOperationsOnMissingData:
+    """Operating on a missing value is an extraction error, not an unhandled raise.
+
+    A single missing lookup returns Undefined, but anything done to that
+    Undefined raises inside Jinja. Callers only handle JinjaExtractionError, so
+    extract() must convert -- while keeping Jinja's message, which names the
+    field that wasn't there.
+    """
+
+    @pytest.mark.parametrize(
+        "expression,data,expected_message",
+        [
+            # A filter or operator that forces the Undefined.
+            ("missing | int", {}, "'missing' is undefined"),
+            ("missing + 1", {}, "'missing' is undefined"),
+            # Reaching into a field the parsed JSON doesn't have.
+            (
+                "(final_message | fromjson).user.status",
+                {"final_message": '{"status": "ok"}'},
+                "'dict object' has no attribute 'user'",
+            ),
+            # Indexing past the end of a list.
+            (
+                "trace[-1].tool_calls[0].function.name",
+                {"trace": []},
+                "has no element -1",
+            ),
+        ],
+    )
+    def test_raising_operation_on_missing_value(
+        self, expression, data, expected_message
+    ):
+        # Only operations that force the value raise; others stay silent, which
+        # test_silent_operation_on_missing_value pins.
+        with pytest.raises(JinjaExtractionError, match=expected_message):
+            extract(expression, data)
+
+    def test_silent_operation_on_missing_value(self):
+        # The other side of the boundary: `~` stringifies Undefined to '' rather
+        # than raising, so a typo'd name yields a wrong value, not an error.
+        assert extract("final_message ~ typo", {"final_message": "hi"}) == "hi"
+
+    def test_lazy_generator_over_missing_field_raises_extraction_error(self):
+        # The sole guard that generators are materialized inside extract()'s try:
+        # map() defers its per-item lookups, so the raise fires during list(),
+        # not when the compiled expression is called.
+        with pytest.raises(
+            JinjaExtractionError, match="'dict object' has no attribute 'tool_calls'"
+        ):
+            extract(
+                "trace | map(attribute='tool_calls.0.function.name')",
+                {"trace": [{"content": "no tools"}]},
+            )
+
+
 class TestTrimAndLstripBlocks:
     def test_trim_blocks_strips_newline_after_block_tag(self):
         template = "{% if True %}\nyes\n{% endif %}"