Rubric Quality pt1: Your Rubric Was Written by Someone Who Has Never Done the Job

If you're doing RL post-training and wondering why your model produces outputs that look right on paper but embarrass anyone who actually works in the domain, the problem might not be the model. It might be the rubric. I see this in closed labs and open source labs alike, and in 2026 it is no longer just my opinion. The papers have caught up.

The afternoon-of-Googling rubric

Here is the pattern. An ML engineer or a data contractor Googles a domain for an afternoon, writes a rubric, and that rubric becomes the standard for what "good" looks like in finance, or law, or insurance. They check for surface-level things: did the model produce an output, does it have the right number of items, is it formatted correctly. They miss the qualitative judgment calls a real professional makes in two seconds.

How big is the gap between an expert and a smart person who Googled it? GPQA measured it. Domain PhDs scored 65% on questions written by their peers. Highly skilled non-experts scored 34%, after averaging more than 30 minutes per question with unrestricted web access.¹ That is the title of this post as a number. If your rubric author is the non-expert, your rubric is wrong about a third of the time before the model ever sees it.

Educational measurement figured this out years before ML did. Susan Brookhart's 2018 paper on rubric criteria attacks exactly the countable-surface-feature rubric: "has three sources" instead of "uses a variety of relevant, credible sources." Her point is that when criteria are countable, the student "need not pay attention to the quality" of the work.² Swap "student" for "policy" and you have this post.

If you're building a rubric to help a chatbot write better legal contracts, does it check that the contract exists and has the expected sections? Or does it check that the indemnification clause actually limits liability in a way a lawyer would consider enforceable, that the governing law matches the jurisdiction, that the payment terms don't create an ambiguity a counterparty would flag on the first read? A first-year associate and a senior partner score the same output very differently, and the senior partner's reasoning is what you want encoded.

The model will optimize for exactly what you reward

This is the part people nod along to and then ignore. If the verifier is informed by surface-level knowledge that nobody in the domain would consider a marker of quality, the model learns to optimize for the wrong things. It produces outputs that look right to someone who doesn't know the domain and embarrassing to someone who does.

A May 2026 paper from the Rubrics-as-Rewards group put a number on it. They name three concrete rubric hacks: partially satisfying a compound criterion, treating implicit content as explicit, and imprecise topical matching. Their headline finding is the one to tattoo on the wall: stronger verification does not prevent reward hacking when the rubric leaves important failure modes unspecified. Rubric-based verifiers preferred the hacked checkpoint while rubric-free judges rated it worse.⁴ A better judge cannot rescue a rubric that never mentioned the thing that matters.

The original Rubrics-as-Rewards ablation makes the provenance point directly. Rubrics generated with expert reference answers scored 35.9 on HealthBench. The same pipeline without expert grounding scored 32.0. The authors' own conclusion: "rubric quality is crucial."⁵

And a frozen rubric is not neutral, it is harmful. Rubric Dropout trained Qwen3-8B with GRPO against a fixed rubric and watched the training reward climb while a stronger gold judge dropped 3 points on HealthBench-Hard and 22 points on ResearchQA. Randomly dropping 30 to 50% of criteria each step, so the policy never optimizes the same rubric twice, recovered most of the loss.⁶ That is the empirical version of what I said in the original post: rubrics and reward functions are evolving artifacts, and the whole post-training team owns them.

Your judge has its own problems

Even a practitioner-written rubric gets scored by something, and that something has biases before it reads your first criterion. Swap the order of two responses and the verdict can flip: with ChatGPT as the evaluator, Vicuna-13B beat ChatGPT on 66 of 80 queries purely by reordering.⁷ Reorder the criteria inside a rubric prompt and the scores move too.⁸ Judges from the same model family as the policy prefer their own outputs, and the effect scales with how well the judge recognizes its own writing.⁹

The most useful 2026 finding is that a consistent judge is not a correct judge. Across 21 judge models, raw agreement with humans overstated chance-corrected agreement by 34 to 41 points: an 85% match rate is roughly a 0.48 kappa. Two production judges had test-retest reliability above 0.95 while carrying severe position bias.¹⁰ If you report "our judge agrees with experts 85% of the time," report the kappa next to it or nobody should believe you.

"Just use tests" is not a free pass either

The standard escape hatch is "we don't need a rubric, we have unit tests." OpenAI audited 138 SWE-bench Verified problems in February 2026 and found 59.4% had flawed tests that reject functionally correct submissions: 35.5% enforced unstated implementation details (one required a function named get_annotation that the problem never mentioned) and 18.8% checked functionality beyond the stated problem.¹¹ This is the benchmark that 93 developers had already filtered by hand, throwing out 68.3% of the original set.¹² A test is a rubric written in Python. It inherits every authorship problem above.

Elo scores, POV rubrics, and evolving artifacts

This is why I like getting both Elo scores and practitioner point-of-view rubrics. They fail in different directions. Under distractor features that a judge happens to like, pairwise preferences flipped about 35% of the time versus 9% for absolute rubric scores.¹³ Meanwhile pure Elo silently rewards formatting: once LMSYS controlled for style, length was the dominant factor and GPT-4o-mini fell from rank 6 to 11.¹⁴ Run both, and treat the disagreements as the interesting data. The pair gives a post-training team more things to reason about when analyzing trends across task outcomes.

The starting point still has to be directionally correct, and directionally correct only comes from people who have lived in the domain. Then the rubric has to keep moving. Shankar and colleagues call it criteria drift: you need criteria to grade outputs, but grading outputs is how you discover your criteria.¹⁵ Some criteria only exist once you have seen a specific failure.

Give the rubric, then explain how the rubric was created

That is the difference between a checklist and a useful artifact. The best published process I know is HealthBench Professional: physicians author the criteria, other physicians peer review them, then an adjudication pass removes ambiguity, explicitly so that low scores are attributable to the model rather than to annotation noise.¹⁶ PaperBench co-developed every rubric with the authors of the paper being replicated and validated the automated judge as its own artifact.¹⁷ Even the original HealthBench, 262 physicians and 48,562 criteria, has a validated tier of 34 consensus criteria and a much larger tier written by one unreviewed physician each.¹⁸ Ask which tier yours is.

Anthropic's agent-eval guidance gives the one-line acceptance test: a good task is one where two domain experts would independently reach the same pass/fail verdict.¹⁹ Run that on 20 items before you run it on 20,000.

Interactive guide: sort the criterion

Checklist: audit your rubric, critic, or verifier

References