Skip to content
New on arXiv: a judge's uncertainty decomposes, so expert labels go exactly where they remove error
Papers and open benchmarks

Research

We build evaluation models, so we spend most of our time on a hard question: how do you know an evaluator is right? These are our own experiments and findings — judge reliability, reward model benchmarks, and the failure patterns we see inside production AI systems.

We publish the negative results too — the techniques that didn't work go out alongside the ones that did. Code and data are open where the underlying data allows it. Four papers, two open benchmarks, a failure-mode ontology, and more on the way.

Papers

  1. arXiv:2609.06444

    Preprint

    Decomposing LLM-Judge Uncertainty to Target Expert Labels

    Ryan Lail ·

    A judge's uncertainty splits in two: the genuine disagreement your experts would have anyway, and the judge's own ignorance of where the bar sits - which labels can fix. A small Bayesian model trained on the labels you already collect separates the two in closed form, with no extra judge calls. Targeting the fixable half removed 83% more error per expert label than escalating on total uncertainty - and a judge's stated confidence, for comparison, barely predicts its errors at all.

    Figure: as expert labels arrive, the judge's epistemic uncertainty drains toward zero while the aleatoric estimate settles onto the true human disagreement level; targeting the epistemic half removes 83% more error per expert label, and stated confidence averages 90.9 out of 100 with no information behind it.
    Read the paper → Code

  2. arXiv:2608.31017

    Preprint

    One note in three: a verified census of three deployed AI scribes, and the instrument that counted it

    Sebastian Fox, Luke Markham, Ryan Lail, Michael Karotsieris ·

    We audited three commercial AI scribes on the same 142 consultations - 565 notes from recorded UK and US encounters plus authored scenarios. Every candidate error went through an adversarial two-model panel told to refute what it could, then blinded clinician adjudication. One note in three carries a verified failure, concentrated in allergy and medication information, invented patient identity, and history written up as examination on telephone consultations that can contain none.

    Figure: 618 verified failures across 565 clinical notes, stacked by the seventeen discovered kinds and coloured by severity, with omissions dominating the count.
    Read the paper → Code Dataset

  3. arXiv:2608.31016

    Preprint

    LLM Judges Verify Presence, Not Absence: Omission Blindness in AI Clinical Notes and What Recovers It

    Sebastian Fox, Luke Markham, Ryan Lail, Michael Karotsieris ·

    The dominant AI scribe error is omission, and LLM judges barely detect it: across eight judge designs, paired discrimination reads 0.79-0.94 on added or altered content but only 0.50-0.63 on omissions, where 0.5 is a coin flip. Prompt wording, voting and GEPA optimisation move the operating point without creating usable detection. Restructuring the task recovers it - list the facts the transcript establishes, then check the note for each.

    Figure: across eight judge designs, paired discrimination reads 0.79 to 0.94 on added or altered content but only 0.50 to 0.63 on omissions, where 0.5 is a coin flip; restructuring the task recovers detection.
    Read the paper → Code Dataset

  4. arXiv:2604.13717

    ICML 2026 workshop paper

    On Cost-Effective LLM-as-a-Judge Improvement Techniques

    Ryan Lail, Luke Markham ·

    Accepted at two ICML 2026 workshops — Statistical Frameworks for Uncertainty in Agentic Systems, and Combining Theory and Benchmarks — Seoul, July 2026.

    We tested four drop-in techniques for improving LLM-as-a-judge accuracy on RewardBench 2. Two of them — task-specific criteria injection and ensembling — reach up to 85.8% accuracy, 13.5 points above a 71.7% baseline, at 1.3x baseline cost, with the gains holding across both the OpenAI GPT and Anthropic Claude families. No fine-tuning required. The techniques that didn't win are published alongside the ones that did.

    Figure: judge accuracy on RewardBench 2 against cost per example - criteria injection plus ensembling reaches 85.8% at 1.3x baseline cost, near the cheapest point on the frontier.
    Read the paper → Write-up Code

In the pipeline

What we're writing up now. Preprints are added above the day they go public.

  • In progress

    What happens to a judge's validity when you optimise it for agreement

Write-ups

Inside 847 Production Clinical AI Notes

Seb's talk at the AI Engineer World's Fair: what we found auditing notes from deployed AI scribes, why an excellent LLM judge shares the scribes' blind spot, and the loop that recovers it.

Seb Fox

A Living Map of LLM Failure Modes: Dominant, Emerging, Squashed

Most failure-mode analysis treats failures as a static taxonomy. In production they're a living distribution - some dominant, some emerging, some being squashed by recent fixes. Here's a clustering pipeline that tracks all three.

Luke Markham

An Ontology of LLM Failure Modes

A structured taxonomy of 60+ failure modes across eight categories, synthesizing recent research into a practical framework for understanding how and why large language models fail.

Luke Markham

LLMs: Great Witnesses, Terrible Judges

LLM-as-a-judge consistency is largely illusory. The same hallucination produces wildly different scores depending on scale configuration, undermining evaluation trust.

Ryan Lail

Working with frontier teams

Everything on this page comes from the same small team, working with domain experts - clinicians first, and specialists in other fields where quality is a matter of judgement. The craft is the instrument: certain-correct answer keys, single-error task design, adversarial filtering before a human ever grades.

If you work on human data, post-training or evals and want expert-graded data, an independent benchmark, or evaluation on tasks where the right answer takes an expert to recognise, we would like to hear from you.

Talk to the founders

Looking for practical guides instead? Read the blog.