Technical Deep-Dive

How JudgeLens Works

From your argument to a data-grounded response in each justice's authentic oral argument style — every step transparent.

The Pipeline

Six steps from your argument to a grounded response. Click any step to learn more.

1
2
3
4
5
6

Data Scale

Every number represents real Supreme Court oral argument data.

43,497+
Exchanges Indexed
Across all 9 justices
1,437+
Oral Arguments
From the Oyez Project
119,849+
Speaking Turns
Justice-attributed
35+
Years of Data
1981–2025

Model Architecture

Five components working together — no single model does all the work.

Embedding Model

nomic-embed-text-v1.5

8,192 token context window. Converts arguments into 768-dimension vectors for semantic search.

Vector Database

ChromaDB

Stores and indexes 5,387 real oral argument exchanges for fast similarity search.

Reranker

cross-encoder/ms-marco-MiniLM-L-6-v2

Reads each candidate exchange alongside the query to produce accurate relevance scores.

Language Model

Claude Sonnet

Generates Alito-style responses constrained by verified behavioral patterns and real exchanges.

Data Source

Oyez Project

16,951 transcript JSON files from Chicago-Kent College of Law. 8,551 unique recordings.

What the Data Reveals: Justice Alito as a Case Study

Patterns discovered by analyzing one justice's 14,691 speaking turns — not assumed, measured. Each of the 9 justices has their own data-driven profile.

16.9%

of turns start with "Well" — and it never means agreement. It signals a pivot, a challenge, or a reframe.

19.9x

"Suppose" is used 19.9x more than attorneys do. It's his primary tool for constructing vivid, everyday hypotheticals.

48%

of deep exchanges end with a statement, not a question. He uses questions to build toward a conclusion.

All 9 sitting justices have their own data-driven profiles built from this same analytical approach.

See the Pipeline in Action

Present your own legal argument and watch JudgeLens transform it into a data-grounded response.

Open the Simulator