Machine Rubrics

The Story

A care-navigation triage assistant scored 4 out of 5 on every rubric dimension — helpful, complete, polite. It was also missing emergency-room referrals. Chest pain with shortness of breath. Sudden one-sided numbness. Severe abdominal pain in a pregnant patient. The responses were thorough. They were also unsafe.

The team's first instinct was the wrong one: the model needs to be smarter. More medical training data. A more capable base model. Weeks of investigation.

The real problem was three lines missing from the rubric. No automatic fail for misses emergency referral when red-flag symptoms are present. No disqualifier for states high-confidence diagnosis despite explicit uncertainty. No separation between good answer and safe answer. The rubric measured quality. It never measured safety. A polite, thorough, unsafe answer scored 4 out of 5 every time.

NurtureBoss lived a quieter version of the same failure. Their property-management assistant kept booking apartment tours on the wrong days, scheduling walk-ins when offices were closed, confusing time zones for multi-location properties. The rubric asked is this helpful? on a 1–5 scale. The scale could not see the dates were wrong.

Two domains. Two failure modes. One root cause: the rubric could not see what mattered.

The Core Idea

A machine rubric is a structured scoring specification that tells a human reviewer or an AI judge exactly what to look for, how to score it, and what constitutes an automatic failure. It replaces "was this response good?" with a checklist of specific, observable criteria -- each with defined score levels and clear examples.

The distinction matters because "good" is not one thing. In the triage story, the response was good on communication and bad on safety. In a fintech support agent, a response can be good on empathy and bad on policy adherence. In a code review assistant, output can be good on thoroughness and bad on signal. Every one of these failures has the same root cause: the rubric collapsed multiple independent quality dimensions into one overall judgment.

The counter-argument deserves its space. The strongest objection to binary rubrics is that they destroy nuance. A financial summary that's 98% perfect but hallucinates one minor date gets the same "Fail" as a response that's entirely fabricated. Engineers tracking incremental improvement need to see progress -- and a binary metric can't show that the failure rate dropped from 40% to 12%.

The synthesis most production teams converge on: binary for the shipping decision (Layer 1), ordinal for diagnostic insight (Layer 3). The top-line tells you whether to ship. The sub-scores tell you where to invest engineering time. The written critique tells you why. You're not choosing between binary and scales. You're using both -- at different layers, for different purposes.

John Snow Labs learned what happens when the rubric measures the wrong dimension entirely. Their Guidelines Central platform matched patient cases to clinical guidelines. The model generated fluent clinical suggestions, but failed to map them to specific SNOMED CT terminologies that EHR systems require. The rubric measured narrative fluency. The clinical workflow needed structured terminology accuracy. The AI was functionally useless until the rubric was redesigned.

Where This Hits in Production

The rubric IS the PRD. In 2025, Zapier's CTO made a cultural shift: the evaluation scoring function itself became the product requirements document. The automatic fail clauses are the hard requirements. The sub-criteria are the quality attributes. The score levels are the acceptance criteria. When the rubric becomes the PRD, every stakeholder is looking at the same artifact.

Multi-tenant rubric metadata. The same response can be acceptable for one customer and a violation for another -- different policies, contracts, risk tiers, geographies. In B2B SaaS, rubrics must be dynamically injected based on the tenant at runtime.

Regulated domains need domain experts writing the fail clauses. Generic quality language -- "accurate," "complete," "helpful" -- is where regulated rubrics go to die. HealthBench's strongest contribution is that 262 physicians authored conversation-specific criteria. A PM can structure the rubric. Only a clinician can write "misses emergency referral when red-flag symptoms present."

Rubrics become release artifacts. In mature teams, the rubric is versioned alongside the prompt and the golden dataset. It has a changelog. It has a calibration set. The rubric is not an annotation memo -- it's part of the release package.

In Practice

The best way to understand machine rubrics is to see three of them side by side -- each shaped by its domain's specific failure modes.

Remember This

1. Four layers, every time: pass/fail gate, automatic-fail clauses, ordinal sub-criteria, written critique. Skip any layer and you lose either the shipping decision or the ability to improve.

2. The most dangerous rubric failure is not a missing criterion. It is letting a 4 on tone average away a 0 on safety. Automatic fails exist to break that arithmetic.

3. Track the unclassified-failure share, not just the pass rate. Pass rate up while unclassified share is also up means your rubric is drifting behind reality — the dashboard is lying to you.