Arc Radius

ArcRadius is a data-driven platform that helps LGBTQ+ young adults understand how legislation affects their rights and provides tools to respond. It integrates classification, explanation, and action generation to translate complex policy into clear, user-centered outputs.

Problem & Motivation

LGBTQ+ young adults in the United States are navigating a rapidly evolving and increasingly hostile policy landscape without accessible tools to understand how legislative changes affect their rights and daily lives. Recent national survey data indicates that political and policy environments are a significant contributor to negative mental health outcomes for this population, with 90 percent of LGBTQ+ young people reporting adverse well being impacts linked to recent political developments (The Trevor Project, 2024).

At the same time, legislative activity targeting LGBTQ+ communities has accelerated at an unprecedented rate. In 2025 alone, more than 600 anti-LGBTQ+ bills were introduced across state legislatures, the highest number recorded in a single year (American Civil Liberties Union, 2025). These bills span critical domains including healthcare access, education, public accommodations, identity documentation, and participation in public life. While information about these policies exists, it is highly fragmented, written in technical legal language, and difficult to interpret without policy expertise.

For affected young adults, this challenge extends beyond information access. It is also a problem of trust. Fragmented sources, inconsistent interpretations, and unclear implications make it difficult for users to determine what information is reliable, whether a policy affects them personally, and what actions, if any, are appropriate. Arc Radius was created to address this gap by helping users move from confusion and uncertainty to understanding and informed civic action.

Designing for Trust

Arc Radius is designed as a trust critical system for a vulnerable user population. Early user research highlighted that credibility, clarity, and perceived care were as important as functional accuracy. As a result, trust served as a central design constraint across both product and technical decisions.

Trust in Arc Radius is operationalized through three core principles:

• Accuracy: All outputs are grounded in authoritative legislative data and validated model pipelines.
• Consistency: Similar inputs produce stable, predictable outputs across the system.
• Clarity: Explanations are written in plain language and do not require policy or legal expertise to interpret.

These principles informed the system architecture, modeling choices, and evaluation strategy, ensuring that users feel informed, supported, and confident in the information they receive.

System Overview

Arc Radius is structured around three core user facing capabilities:

1. Understand: Identify legislation that is relevant to the user.
2. Interpret: Explain what the legislation does and why it matters.
3. Act: Enable users to respond through concrete civic actions.

These capabilities map directly to three system components: a classification pipeline, GraphRAG based explanation generator, and an action generation module.

Data Sources & Approach

Arc Radius integrates large-scale legislative data with curated labeled datasets.

Primary data sources include:

• LegiScan API for bill text, metadata, and legislative history
• Labeled datasets from ACLU, Plural, and related organizations

The full dataset contains approximately 750,000 bills from 2021 through 2026. Of these, 2,325 bills are labeled as LGBTQ+ related. To address severe class imbalance, a 2:1 negative sampling strategy was used, resulting in an additional 4,650 non LGBTQ+ bills included in the training set.

Modeling Constraints

Several constraints shaped the modeling approach:

• Limited labeled data for LGBTQ+ specific legislation.
• Extreme class imbalance between relevant and non relevant bills.
• Ambiguity and domain specificity of legal language.

These constraints motivated the use of a two stage classification pipeline and the incorporation of metadata features for stance prediction, rather than relying on a single end to end text model.

System Architecture

The system follows a modular pipeline architecture. Each stage is decoupled from the others.

Pipeline stages:

1. Data ingestion and normalization
2. Relevance classification
3. Stance classification
4. Graph based retrieval and generation
5. Letter generation

Each stage operates independently and produces validated outputs that feed into the next component. This separation reduces error propagation, simplifies debugging and iteration, and supports independent evaluation of system components. From a user perspective, this design improves reliability by ensuring that each layer of interpretation is grounded and validated before being surfaced.

Core System Components

1. Classification: What is this bill?

The classification component determines whether a bill is relevant to LGBTQ+ issues and, if so, whether it is supportive or harmful. This task is implemented as a two stage pipeline to improve interpretability and reduce compounding errors.

Stage 1: Relevance Classification

Relevance classification is performed using a LegalBERT model trained on bill titles and short descriptions. Only text inputs are used at this stage to avoid metadata leakage. Given that errors at this stage would exclude bills entirely from downstream analysis, precision for LGBTQ+ relevant bills is prioritized.

LegalBERT achieved the strongest overall performance among evaluated models, including general purpose transformer and instruction tuned models. Its ability to capture legal language patterns resulted in high accuracy and stable performance across stratified train test splits.

Stage 2: Stance Classification

Stance classification is applied only to bills identified as LGBTQ+ relevant. Multiple approaches were evaluated, including text based transformer models and metadata driven models.

Text based models showed limited improvement due to the small labeled dataset and the inherent ambiguity of legal language. In contrast, metadata driven models performed more consistently. Features such as sponsor party composition, state political lean, and legislative context provided strong and interpretable signals of bill intent.

Based on these findings, a linear model using engineered political context features was selected. Model selection prioritized minimizing false supportive classifications of harmful legislation, as such errors pose higher risk to users.

Classification Model Comparison & Performance

Relevance Model Comparison & Performance