How to Write Reliable Rubrics for LLM-as-a-Judge Evaluations

Follow up to Part 1: How to Design AI Evaluations You Can Actually Trust

At Google, we are publishing a suite of Agent Skills for Google products and technologies on GitHub. My team is interested in measuring their performance to understand how they perform. Deterministic tests, like checking if generated code compiles, are ideal. Unfortunately they can't be easily created at scale for nuanced, generative responses, such as answers to open-ended questions or information retrieval tasks.

In my previous post we looked at what you test, which means the evaluations that are the actions you ask the agent to perform. The next step is to look at how you assert whether the agent succeeded. This means creating reliable and accurate assessments of the agent's response.

To assess complex outputs at scale, especially when topics cover broad areas with nuanced parts, we use an "LLM-as-a-judge" approach. Responses are assessed against a structured rubric using a model based grader. The judge evaluates each response using a set of true/false questions. Aggregated, these answers provide an accuracy score for a response.

Giving an LLM a vague prompt or subjective questions leads to ambiguity in its response. This ambiguity introduces noisy data and causes inconsistent evaluations. Ultimately, it wastes your token budget on metrics that aren't useful.

To make these assessments more reliable, you must treat your rubrics like formal specifications. By constraining the judge to evaluate strict, objective boolean truths, you reduce the chance of hallucinations. Because evaluating strict boolean truths is a less complex task, you can even use smaller, faster models for grading.

Here are four lessons we learned to help you write robust rubric questions for your LLM-as-a-judge scorer.

Keep Questions Atomic and Distinct

Evaluating multiple requirements in a single question, such as "Does the response contain a metadata property and format the output as JSON?", forces the LLM judge to guess which clause is more important. This ambiguity leads to inconsistent grading and wasted tokens.

• Split compound questions: Instead of writing one large check, divide your requirements into discrete, atomic TRUE/FALSE questions. (e.g., Check 1: Does it contain a metadata property? AND Check 2: Is the output JSON?)
• Avoid overlapping questions: Never test the same underlying concept multiple times across your rubric. Overlap risks double-penalizing the evaluated model for a single mistake, which corrupts your accuracy score.
• Reduce the reasoning load: Eliminate the need for the judge to weigh competing clauses. When every question assesses exactly one distinct fact, your grading becomes more consistent.

Constrain the Judge: Objective Facts Over Subjective Reasoning

The rubric-based approach exists because giving an LLM judge a full prose prompt to evaluate a complex response leads to inconsistent numbers. If you ask the judge subjective questions like "Was this a comprehensive answer?" or ask it to explain "Why did the agent do this?", you introduce ambiguity that creates noisy, unrepeatable data.

• Focus on observable facts: Do not ask the judge to evaluate concepts that require interpretation, such as intent, quality, or reasoning. Only assess concrete facts that you expect to find in the response.
• Write formal specifications: Use strict, objective language, such as RFC 2119 terminology (MUST, MUST NOT, REQUIRED), to test for observable outcomes. The judge should never have to guess what you mean.
• Test for negative constraints: Explicitly verify what the agent should not do. Instead of asking if the agent "used best practices", check that it did not suggest a specific deprecated feature.
• Require strict true/false answers: By forcing a strict TRUE/FALSE classification on objective facts, you reduce the reasoning load and reduce the variance in ratings.
• Avoid opportunities for cheating: Agents will tailor answers to game your tests if given the chance. Keep scoring rubrics isolated in a separate system. Design rubrics that focus on strict functional outcomes or specific topics, rather than broad keywords matches.

Grade Only What You Asked For

When building rubrics, it is easy to accidentally evaluate an agent on requirements that were never stated in the prompt. Doing this creates false negatives and reduces the accuracy of your measurement.

• Align the rubric to the prompt: Assess only what was explicitly requested. For example, do not penalize a model for failing to provide citations if the prompt never asked for them.
• Grade the destination, not the journey: Avoid writing rubrics that check if the agent used a specific tool or followed a rigid sequence of steps. Pre-trained models might bypass custom tools entirely if they already know the answer.
• Evaluate the final response: Grade the objective output. If you need to evaluate a step-by-step process, prompt the agent to output an execution plan and evaluate the plan itself.

Calibrate Your Judge

Even if you follow these rules and write perfectly atomic, objective questions, your LLM judge might still misinterpret your scoring instructions and rubrics. To guarantee your pipeline generates consistent ratings and reliable signals, you must prove that the judge's scoring aligns with how a human subject matter expert would evaluate the exact same response through calibration.

• Establish a human baseline: Ask subject matter experts to manually grade a "golden set" of test responses.
• Run the comparison: Run your LLM-as-a-judge against this golden set and compare the automated scores against the human scores.
• Identify the delta: If the LLM judge disagrees with your human experts, it is usually a sign that your rubric or the grading instructions are too ambiguous.
• Iterate until aligned: Tweak and calibrate your rubric questions or the scoring instructions until the LLM judge consistently aligns with human experts. Only then is your judge ready.

Summary

Once you have this reliable data, the next step is making it visible. In AI Evals at a Glance: Heatmaps for Stakeholders Joe Spiro explains how to take these raw measurements and visualize evaluations.

When building our agent skills, we learned that ambiguous evaluation rubrics do not provide useful signals and feedback. Forcing your LLM judge to evaluate strict, boolean facts eliminates this noise. It makes your tests repeatable, optimizes your token spend, and lets you confidently measure whether your AI tool is actually improving.

Photo by William Warby on Unsplash