Symio-ai/legal-evidence-classifier

Model Description

Legal Evidence Classifier classifies evidence by type, admissibility, relevance, and evidentiary weight. Given a piece of evidence (document text, testimony excerpt, or exhibit description), it predicts: evidence type (documentary, testimonial, physical, digital), admissibility under FRE/state evidence rules, likely objections, and probative value for each cause of action.

Critical for exhibit management and evidence strategy in the GLACIER pipeline.

Intended Use

• Primary: Classify and organize evidence for trial preparation and filing
• Secondary: Predict admissibility challenges and suggest foundations for admission
• Integration: Powers exhibit management in GLACIER Stages 4 and 5

Task Type

text-classification -- Multi-label evidence classification with admissibility scoring

Base Model

microsoft/deberta-v3-large -- Strong NLI for evidence classification and admissibility analysis

Training Data

Source	Records	Description
Evidence Rulings	~300K	Judicial rulings on evidence admissibility
FRE Case Annotations	~100K	Federal Rules of Evidence case annotations
FL Evidence Code Cases	~50K	Florida evidence code rulings
Exhibit Lists (with outcomes)	~200K	Exhibit lists from trials with admission/exclusion results
Expert Witness Challenges	~50K	Daubert/Frye challenge outcomes

Classification Categories

• Type: Documentary, testimonial, physical, digital, demonstrative, expert
• Admissibility: Admissible, likely objected, likely excluded, privileged
• FRE Issues: Hearsay (with exception analysis), authentication, best evidence, relevance (403 balancing)
• Foundation: What foundation is needed for admission
• Weight: High/medium/low probative value per cause of action

Common Objection Predictions

• Hearsay (FRE 801-807) with exception identification
• Authentication (FRE 901-902)
• Best evidence (FRE 1001-1008)
• Relevance/prejudice (FRE 401-403)
• Privilege (attorney-client, work product, spousal)
• Expert reliability (Daubert/Frye)

Benchmark Criteria (90%+ Target)

Metric	Target	Description
Type Classification	>= 95%	Correct evidence type
Admissibility Prediction	>= 82%	Matches actual judicial ruling
Hearsay Exception ID	>= 88%	Correctly identify applicable hearsay exception
Objection Prediction Recall	>= 85%	Predict objections opposing counsel will raise
Foundation Requirements	>= 90%	Correctly specify needed foundation

GLACIER Pipeline Integration

STAGE 2 (Research) --> evidence-classifier organizes available evidence
STAGE 4 (Draft) --> classifier determines exhibit order and index
STAGE 5 (WDC #2) --> verify all exhibits have proper foundation and no admissibility gaps

Training Configuration

• Epochs: 8
• Learning rate: 2e-5
• Batch size: 16
• Max sequence length: 1024
• Hardware: AWS SageMaker ml.g5.4xlarge

Limitations

• Admissibility is highly context-dependent; prediction is probabilistic
• Judge-specific evidence rulings are not modeled (see judicial-predictor for that)
• Cannot analyze physical evidence from description alone
• Digital evidence authentication requirements evolve rapidly
• Does not perform chain-of-custody analysis

Version History

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