Evaluate

KLRB

A Kubernetes benchmark that measures whether an LLM actually read the cluster evidence before diagnosing an incident — or just guessed confidently from metadata.

★ 96 on GitHub Python · Kubernetes · Chaos engineering · Multi-provider LLM runner

Try the ablation

One Kubernetes incident, three conditions. Take away the evidence that justifies the model's answer and watch what the answer does. This takes about ten seconds and it is the entire finding.

Evidence handed to the model 6 of 6
    The model's root cause

    stated confidence 0.91

    A worked example of the ablation the harness runs. The three conditions and the evidence shape are the real ones; the response text is representative of the OOMKill scenario rather than a verbatim published run.

    Why this measurement and not accuracy

    You can score an incident responder on whether its answer matched the label. That number looks fine and tells you almost nothing, because on a corpus of common Kubernetes failures a model can reach it from priors alone — CrashLoopBackOff on a pod with a low memory limit is an OOMKill often enough that guessing pays.

    An accuracy score cannot separate a responder that read your cluster from one that recognised the shape of your pod name. In production that difference is everything: the first one generalises to your weird outage, the second one is confidently wrong precisely when you need it most.

    So KLRB scores the dependency instead of the answer. Run the scenario with the causal evidence, run it without, and compare. If the conclusion survives the removal of its own justification, the justification was decorative. I call the failure mode the Confident Liar effect — not because the model intends to mislead, but because fluency and groundedness turn out to be independent, and only one of them is visible in the output.

    How it is built

    Scope — what this is not

    A benchmark harness and a reproducible effect, not a peer-reviewed result. Scores are comparable within the harness. I would not yet quote an absolute number as a property of a model in general — the honest claim is the direction and the method.

    The open questions I would want a reviewer to press on: scenario count and seeds per condition, whether the ablation control removes the evidence or merely corrupts it (both are implemented, they measure different things), and how much of the effect is the model versus my prompt's log formatting. A preprint is in progress and these are the sections that matter.

    Elsewhere in the system