When several imperfect judges label the same items, the disagreement pattern itself reveals who is reliable. judgepanel fits the classic Dawid–Skene model (1979) to a judge panel and recovers, jointly:

• each judge's confusion matrix (sensitivity/specificity in the binary case),
• the class prevalence,
• a posterior over every item's true label, weighted by estimated judge quality instead of one-judge-one-vote.

Around the core: majority-vote baselines with explicit tie handling, chance-corrected agreement (Cohen's and Fleiss' kappa), bootstrap confidence intervals, and simulators for planning panels and studying failure modes. Pure NumPy; no model calls; every fit deterministic.

Install

pip install judgepanel

Sixty seconds

Your data is one JSON object per (item, judge) verdict:

{"item": "case-0017", "judge": "gpt-strict", "label": "faulty"}

judgepanel fit examples/data/demo_panel.jsonl --positive faulty --bootstrap 200

panel: 300 items, 5 judges, 2 classes (clean, faulty)
converged after 55 iterations
estimated prevalence: clean=0.577, faulty=0.423

judge            sens    spec  youden
gpt-strict      0.906   0.931   0.836
gpt-lenient     0.715   0.999   0.714
small-model     0.692   0.710   0.402
rules           0.562   0.996   0.559
flag-happy      0.953   0.394   0.347

The demo panel is synthetic with known truth (examples/data/demo_gold.jsonl): the true parameters are (0.92, 0.94), (0.70, 0.97), (0.66, 0.72), (0.55, 0.99) and (0.97, 0.35) — every estimate above lands within 0.04 without ever seeing a gold label, and the flag-everything judge is exposed immediately.

Same thing in Python:

from judgepanel import fit_dawid_skene, read_panel_jsonl

panel = read_panel_jsonl("panel.jsonl")
result = fit_dawid_skene(panel)
result.sensitivity("faulty")   # {"gpt-strict": 0.906, ...}
result.labels()                # posterior-aggregated label per item

Keys are configurable, so existing verdict files work directly (--item-key trajectory_id --judge-key judge_id --label-key faulty; boolean labels become "true"/"false").

Validation on a real judge panel

The trajectory-judge study had five LLM judges label the same 400 agent trajectories as faulty or clean, with gold labels known by construction. Fitting judgepanel on the verdicts alone (gold hidden), then unblinding (python examples/case_study.py --raw .../results/raw):

Three honest findings:

1. The spammer is exposed exactly. The 8B judge flags nearly everything; its estimated specificity is 0.000 — matching the truth — with no gold label in sight. The judge quality ranking is right across the board, and the production-default outcome judge is estimated within 0.04.
2. Correlated judges inflate their own estimates. Self-consistency is the step judge sampled three times, and the two agree at kappa 0.96. Dawid–Skene assumes conditionally independent judges, so this pair drags the latent truth toward its shared verdicts: both get overestimated by up to +0.14, and estimated prevalence drops to 0.645 against a true 0.750. The library prints the kappa matrix precisely so you check this before believing the fit.
3. No aggregation recovers what every judge misses. Faults that leave the final answer correct are invisible to the whole panel, so posterior aggregation (0.895 accuracy) cannot beat majority (0.897) here. Weighted aggregation wins when judge quality is uneven, not when errors are shared.

When not to trust it

• Fewer than 3 judges: weakly identified; the fit warns.
• Highly correlated judges (same base model, ensembles of one judge, shared blind spots): estimates bias toward the correlated cluster — check judgepanel agree first, and prefer diverse judge families.
• A panel that is wrong on average: if most judges are worse than chance, the labels-and-mirror solution is indistinguishable without gold.
• Tiny or extremely imbalanced panels: run judgepanel simulate with your expected operating points to see the estimation error you should expect before you rely on the numbers.

API

CLI: judgepanel fit, judgepanel agree, judgepanel simulate, judgepanel version (each supports --json).

Development

python -m venv .venv && .venv/bin/pip install -e ".[dev]"
.venv/bin/pytest          # 52 tests: hand-computed values, seeded recovery, CLI
.venv/bin/ruff check src tests && .venv/bin/mypy src

The suite checks statistical behaviour, not just plumbing: kappa against hand-computed tables, EM recovery on seeded synthetic panels, log-likelihood monotonicity, bootstrap coverage, and determinism.

Part of evalstack

evalstack is the map of these eleven evaluation tools: what each one measures, what it found on real data, and the two chains that run end to end.

Citing

The accompanying paper, Judging Without Gold: Gold-Free Estimation of LLM-Judge Accuracy and Its Failure Modes, is a working paper (2026).

Releases are archived on Zenodo. Cite the concept DOI 10.5281/zenodo.21809473, which always resolves to the latest version; structured metadata is in CITATION.cff.

License

MIT — see LICENSE.