Evaluating judicial decision-support models transparently

What would transparent evaluation of a decision-support model require?

Define the exact support task, intended users, decision consequences, reference standard, test population, and failure costs before reporting a score. Publish the limitations and human review process alongside performance. A benchmark result cannot establish that a model should exercise judicial authority.

Evaluate one task, not an undefined 'AI judge'

A model that flags potentially missing attachments has a different task from one that recommends a legal outcome. Specify the input, output, excluded uses, and permitted downstream action. NIST's GAI profile calls for documented evaluations tied to deployment conditions and generalization limits. [S1]

• Task: identify whether a defined required item is present in a synthetic packet.
• Output: flag with source location and an uncertainty option.
• Reference: two qualified reviewers' labels, disagreement resolution, and rationale.
• Excluded claim: legal sufficiency, credibility, entitlement to relief, or final adjudication.

Synthetic error matrix: omissions in 100 packets

An invented test set contains 20 packets with missing items and 80 complete packets. The model flags 16 missing-item packets and incorrectly flags 8 complete packets. It misses 4 omissions and correctly clears 72 complete packets. Recall for omissions is 16/20 = 80%; precision of flags is 16/24 ≈ 66.7%; overall accuracy is 88/100 = 88%.

• Accuracy alone hides the four missed omissions.
• Review workload is 24 flagged packets, including eight false alarms.
• A decision threshold should reflect the consequences of each error, not only a higher headline score.
• These invented figures establish no real model capability.

Test what can break the claimed use

Use separate development and evaluation sets, prevent duplicate leakage, document selection criteria, and inspect performance across relevant document conditions. CEPEJ's charter emphasizes quality/security, non-discrimination, and transparent methods; the test design here is a proposed implementation of those concerns. [S2]

• Unreadable scans, missing pages, unusual layouts, language variation, and conflicting metadata.
• Changed forms or procedural requirements after the evaluation date.
• Packets outside scope and cases where the model should abstain.
• Different reviewers' interpretation of an ambiguous reference standard.
• Whether human reviewers catch model mistakes or merely accept its framing.

Publish a release decision with bounded claims

A useful report includes model/version, source data permissions, test dates, task definition, error examples, subgroup or condition results where justified, uncertainty, and the person responsible for release. Preserve unsuccessful runs and the reasons for exclusion. Independent replication should be possible without exposing private case material.

• Report observed performance only for the tested conditions.
• State what remains untested and what requires human determination.
• Define monitoring and suspension triggers before use.
• Reevaluate after a material model, form, data, or workflow change.

Scope and limitations

A proposed evaluation protocol for research. This site is not a judge, court, judicial service, or endorsement of automated adjudication.

No model was tested, no court system was evaluated, and no threshold is recommended for actual judicial use.

Sources

1. NIST AI 600-1 — Generative Artificial Intelligence Profile — GAI risk and evaluation guidance; Measures 2.3/2.5 address capability evidence, generalization limits, and source verification. Checked 2026-11-05
2. Council of Europe CEPEJ — European Ethical Charter on AI in judicial systems — 2018 ethical principles for rights, non-discrimination, quality/security, transparency, and user control; not a universal binding rule. Checked 2026-11-05

Explore further

Evaluation design helps frame careful questions about legal-team AI implementation. Explore the broader workflow topic at Agentic Attorney: https://agenticattorney.com/ . Neither destination exercises judicial authority. Explore Agentic Attorney.