Transparency

Methodology & Data Standards

JudgeFinder summarizes public court activity to make judicial decision patterns easier to understand. This page explains how records flow into the platform, how we normalize outcomes, and what powers the confidence scores that appear on every metric.

Data inputs

We combine public court dockets, published decisions, and structured CourtListener feeds to build daily datasets.

•CourtListener API: decisions, dockets, judge assignments, and metadata refreshed twice daily.
•State and county portals: availability checks run nightly with automated retrievers for missing filings.
•Manual reconciliation: compliance team validates judge identity, courtroom, and assignment changes before publishing.

Normalization & deduplication

Court documents use inconsistent labels. We normalize party roles, case outcomes, and judge identifiers before analysis.

•Party roles: plaintiff/defendant or petitioner/respondent mapped to civil or family categories with fallback heuristics.
•Outcome harmonization: raw text is classified into settle, dismiss, judgment, plea, or other buckets with rules that surface ambiguous language for review.
•Duplicate guardrails: docket numbers and filing timestamps are hashed to prevent double counting when multiple feeds overlap.

Confidence scoring

Every metric includes a 60–95% confidence score based on sample size, document clarity, and model agreement.

•Sample size weighting: scores degrade when n < 25 and hide below the public threshold (default 15).
•Document quality: OCR clarity, missing outcome text, or sealed filings reduce confidence in proportion to uncertainty.
•Model agreement: Gemini 1.5 Flash and GPT-4o-mini must agree within 6 percentage points or the fallback routine reruns with stricter prompts.

Limitations & safeguards

AI analytics support transparency but do not replace legal advice.

•Coverage gaps: counties with PDF-only archives may lag up to 14 days; these are labeled LOW quality until verified.
•Bias detection boundaries: we surface statistical tendencies only; we never label a judge as biased or make disciplinary claims.
•Human-in-the-loop: analysts review edge cases weekly, and all model prompts are version-controlled with regression tests.

Sampling thresholds

Metrics are hidden on the public site when the sample size falls below the configured threshold (environment variableNEXT_PUBLIC_MIN_SAMPLE_SIZE, default 15). Scores between 15 and 30 show a LOW quality badge and direct users to request an update. All denominators display next to each chart so readers understand the evidence behind a percentage.

Need more depth?

For details on governance, prohibited uses, or how we label advertisements, explore the rest of the trust library.

Governance & Ethics Advertising policy Release changelog

Data inputs

We combine public court dockets, published decisions, and structured CourtListener feeds to build daily datasets.

•CourtListener API: decisions, dockets, judge assignments, and metadata refreshed twice daily.

•State and county portals: availability checks run nightly with automated retrievers for missing filings.

•Manual reconciliation: compliance team validates judge identity, courtroom, and assignment changes before publishing.

Normalization & deduplication

Court documents use inconsistent labels. We normalize party roles, case outcomes, and judge identifiers before analysis.

•Party roles: plaintiff/defendant or petitioner/respondent mapped to civil or family categories with fallback heuristics.

•Outcome harmonization: raw text is classified into settle, dismiss, judgment, plea, or other buckets with rules that surface ambiguous language for review.

•Duplicate guardrails: docket numbers and filing timestamps are hashed to prevent double counting when multiple feeds overlap.

Confidence scoring

Every metric includes a 60–95% confidence score based on sample size, document clarity, and model agreement.

•Sample size weighting: scores degrade when n < 25 and hide below the public threshold (default 15).

•Document quality: OCR clarity, missing outcome text, or sealed filings reduce confidence in proportion to uncertainty.

•Model agreement: Gemini 1.5 Flash and GPT-4o-mini must agree within 6 percentage points or the fallback routine reruns with stricter prompts.

Limitations & safeguards

AI analytics support transparency but do not replace legal advice.

•Coverage gaps: counties with PDF-only archives may lag up to 14 days; these are labeled LOW quality until verified.

•Bias detection boundaries: we surface statistical tendencies only; we never label a judge as biased or make disciplinary claims.

•Human-in-the-loop: analysts review edge cases weekly, and all model prompts are version-controlled with regression tests.

Sampling thresholds