Prior art & method citations¶
muteval builds on established statistics and a body of eval/meta-evaluation
research. This page names the lineage — both to give credit and to preempt the
fair questions "isn't this just mutation testing?" and "isn't judge-reliability
already solved?". Short answer: the mutation-testing-for-eval-suites framing and
its absence detection are novel as a product; the probes port well-known
statistics and, for judge reliability, reach parity with an already-served area.
(Full competitive audit: docs/AUDIT-probe-prior-art.md.)
The core idea (mutation testing for evals)¶
- MILE — Wei et al., Mutation Testing of In-Context-Learning Systems (arXiv 2409.04831, SETTA 2024). The closest prior work and an independent confirmation of muteval's central claim that a mutation score tracks eval-suite quality. Scoped to ICL classification, a research prototype — muteval generalizes it (prompt/context/tools/model), packages it (CLI/adapters/survivor abstraction), and adds absence detection.
- DeepMutation (arXiv 1805.05206) and the broader ML-mutation-testing line mutate models to grade traditional-ML test data — adjacent, different axis.
mutmut/ Stryker — mutation testing for code; the "mutation score = test suite quality" heritage muteval transposes to evals.- Tangled up in BLEU — Mathur et al., ACL 2020 (aclanthology 2020.acl-main.448). Metric-reliability critique; its Type-I/II framing maps onto muteval survivors as Type-II errors ("your suite would miss this").
- CheckList — Ribeiro et al., ACL 2020 (2020.acl-main.442). Software-testing discipline for NLP; tests models, not suites — a philosophical ancestor muteval can consume rather than compete with.
Probe method lineage¶
statistical_adequacy — Wilson score interval (default) + Jeffreys/Beta-Binomial
for small n.
- Brown, Cai & DasGupta (2001), Interval Estimation for a Binomial Proportion —
the binomial-CI authority (use Wilson/Jeffreys over Wald).
- Bowyer, Aitchison & Ivanova (2025), Don't Use the CLT in LLM Evals With Fewer
Than a Few Hundred Datapoints (arXiv 2503.01747; lib bayes_evals) — argues
muteval's exact thesis; the reason Jeffreys is offered for tiny n.
- Evan Miller (2024), Adding Error Bars to Evals (arXiv 2411.00640) — the
power / "how many cases" framing.
- Card et al. (2020), With Little Power Comes Great Responsibility (EMNLP) — the
canonical "your eval is too small" result.
judge_reliability — verdict-flip rate + Krippendorff's alpha over re-runs. - Reliability without Validity (arXiv 2606.19544) — large-scale study using Krippendorff's alpha for judge test-retest; source of the reliability≠validity caveat muteval states. - The Coin Flip Judge (arXiv 2606.13685) — ~11 runs for a stable majority; temperature sensitivity (why the probe recommends temp 0). - Zheng et al. (2023), Judging LLM-as-a-Judge with MT-Bench (arXiv 2306.05685) and Wang et al., LLMs are not Fair Evaluators — the bias taxonomy (position/ verbosity/self-preference) and the CALM framework (Justice or Prejudice, ICLR 2025) are the blueprint for the directional-bias panel (future work; needs a structured judge abstraction). - Honest note: judge alignment/reliability is already served (LangSmith Align Evals, Ragas, Arize, standalone MetaEvaluator). muteval's probe is parity, not a differentiator.
discrimination — AUC (Mann–Whitney U) + Cohen's d + significance. - Classical Test Theory item-discrimination index (point-biserial); recent LLM application with a ~0.15 threshold (arXiv 2606.18709). - Kocmi et al., To Ship or Not to Ship (WMT21, arXiv 2107.10821) and Ties Matter (arXiv 2305.14324) — the NLG/MT field's move to rank/pairwise measures over raw score gaps, which is why muteval uses AUC not a mean gap.
redundancy — Spearman rank correlation + connected-component families. - scikit-learn's multicollinearity recipe (Spearman + hierarchical clustering). - Agreement Metrics for LLM-as-Judge (arXiv 2606.00093) — reporting several correlated metrics is an "illusion of corroborating evidence" (the probe's thesis).
Neighbours in the eval-quality space (not the same axis)¶
- MetaEvaluator (GovTech) — rates judges (Cohen's kappa, alt-test), the nearest standalone package; overlaps the probe layer, not the mutation core.
- FBI (AI4Bharat, EMNLP 2024) — perturbs inputs to find evaluator-LLM blind spots; closest mechanism to muteval's, judge-scoped, academic.
- Incumbent eval-quality features (LangSmith Align Evals, Ragas Align, Arize meta-evals, DeepEval metric-alignment) are almost all judge-vs-human alignment — which by construction cannot surface "you have no eval for this at all."
Positioning (evidence-backed)¶
The moat is the mutation core + absence detection and, secondarily, discrimination and redundancy (essentially unserved by any tool). judge_reliability is table-stakes parity — never lead with it. The eval-evaluator superset stays demand-gated.