Legal Judicial Predictor predicts how a specific judge will rule on a given motion or legal issue. Given a judge identifier, motion type, legal issue, and case facts, it predicts the likely outcome (granted/denied/partial) with a confidence score, based on the judge's historical ruling patterns.

Provides strategic intelligence for the GLACIER WDC panels on judge-specific tendencies.

Intended Use

Primary: Predict likely judicial outcomes to inform litigation strategy
Secondary: Guide motion practice (which motions to file, which to avoid)
Integration: Feeds into GLACIER Stage 3 WDC for strategy validation

Task Type

text-classification -- Multi-class prediction (granted/denied/partial/deferred) with confidence

Base Model

nlpaueb/legal-bert-base-uncased -- Legal domain pre-training for understanding judicial reasoning patterns

Training Data

Source	Records	Description
CourtListener Judge Profiles	~35K judges	Judge biographical and appointment data
Docket Entry Rulings	~5M	Motion outcomes by judge, type, and issue
Published Opinions	~2M	Full-text opinions showing reasoning patterns
Oral Argument Transcripts	~100K	Judicial questioning patterns
Sentencing Data	~500K	Sentencing patterns (for criminal cross-reference)

Prediction Features

Judge identity: Historical ruling patterns by motion type
Motion type: MTD, MSJ, MIL, sanctions, discovery, etc.
Legal issue: Specific legal question presented
Party characteristics: Pro se vs. represented, plaintiff vs. defendant
Case type: Tort, contract, statutory, constitutional
Jurisdiction context: Circuit/district patterns

Output Format

json

{
  "judge": "Judge Smith",
  "motion_type": "Motion to Dismiss (12(b)(6))",
  "prediction": "DENIED",
  "confidence": 0.73,
  "judge_historical_rate": "Grants 35% of 12(b)(6) motions",
  "key_factors": ["judge favors liberal pleading standard", "similar issue denied in Case X"],
  "comparable_rulings": ["Case A (denied)", "Case B (denied)", "Case C (granted)"]
}

Benchmark Criteria (90%+ Target)

Metric	Target	Description
Prediction Accuracy	>= 72%	Overall outcome prediction (judicial behavior is inherently uncertain)
High-Confidence Accuracy	>= 88%	Accuracy when confidence > 0.8
MTD Prediction	>= 75%	Motion to dismiss outcome prediction
MSJ Prediction	>= 70%	Summary judgment outcome prediction
Calibration Error	<= 0.05	Predicted confidence matches actual accuracy

Note: 90%+ target applies to the model's calibration and high-confidence predictions, not raw accuracy -- judicial behavior has inherent unpredictability.

GLACIER Pipeline Integration

STAGE 1 (Classify) --> identify assigned judge
STAGE 3 (WDC #1) --> judicial-predictor informs strategy (which motions to file)
STAGE 5 (WDC #2) --> predictor validates the filing strategy against judge tendencies

Training Configuration

Epochs: 10
Learning rate: 2e-5
Batch size: 32
Max sequence length: 512
Hardware: AWS SageMaker ml.g5.2xlarge

Limitations

Judicial prediction is inherently uncertain; model provides probabilities, not certainties
New judges with limited history will have low-confidence predictions
Judges change behavior over time; recency weighting helps but doesn't eliminate drift
Does not account for en banc or appellate override dynamics
Ethical consideration: should inform strategy, never be the sole basis for litigation decisions

Version History

Version	Date	Notes
v0.1	2026-04-10	Initial model card, repo created