A metric created with Metric.create(...) and used through measure() sends an empty string to the judge model, so the resulting scores are judgments of nothing. There is no error — the run completes and returns numbers that look valid.

Verified against aixplain==0.2.48 (Python 3.13) by introspection and local execution. No network calls needed to reproduce the core of it.

The chain

1. Metric.agent_response_data_fields defaults to AgentResponseDataFields(query=False, trace=False, output=False) — all three flags off.
2. Metric.create(...) never sets agent_response_data_fields (the name does not appear in its body).
3. measure() builds its payload from those flags:

def measure(self, agent_response):
    metric_input = self.agent_response_data_fields.give_metric_input(agent_response)
    if self.additional_input_prompt:
        metric_input = self.additional_input_prompt + "\n\n" + metric_input
    metric_payload = {"data": metric_input}
    return self.run(data=metric_payload)

4. give_metric_input() appends only the enabled fields, so with all three False it returns "".

Reproduction (no API key required)

from aixplain.v2.agent_evaluator import AgentResponseDataFields

class R:  # stand-in for AgentResponseData
    input = "What is 2+2?"; output = "4"; steps = []

print(repr(AgentResponseDataFields().give_metric_input(R())))         # ''
print(repr(AgentResponseDataFields(output=True).give_metric_input(R())))  # 'Output: 4\n'

The first line is what a Metric.create(...) metric sends as {"data": ""}.

Impact

Any evaluation built on Metric.create + measure() silently scores an empty prompt unless the caller knows to set agent_response_data_fields themselves. Because the call succeeds and returns a score, there is nothing to alert the user that the judge never saw the agent's output.

Suggested fixes (any one would do)

• Default AgentResponseDataFields to output=True (the minimum useful payload), or
• have Metric.create(...) set agent_response_data_fields from its arguments, or
• raise in measure() when the resulting metric_input is empty.

Secondary, same method

Metric.create(...) begins with del allowed_actions, kwargs — the allowed_actions parameter is accepted and silently discarded. Either honour it or drop it from the signature.

Related note: the Eval docstring states each metric is invoked with a payload containing output and reference; in the current code EvalCase.reference never reaches the metric at all.