Hubble is an agentic framework for automated alpha factor discovery that integrates large language models (LLMs) with a domain-specific operator language, an abstract syntax tree (AST) execution sandbox, a dual-channel retrieval-augmented generation (RAG) module, and a family-aware selection mechanism. The system restricts LLM generation to interpretable operator trees, evaluates candidates through a deterministic cross-sectional pipeline, and feeds back top formulas and family-level diagnostics to subsequent rounds. The dual-channel RAG provides positive references to encourage exploration of under-covered themes and negative references to discourage crowded templates. The scoring layer standardizes multiple metrics (RankIC, Pearson IC, turnover, etc.) and applies penalties for crowding, similarity, and family concentration, along with a novelty bonus, ensuring diversity in the final selection.

In a main experiment on a U.S. equity universe of approximately 500 stocks (840 trading days from 2022-01-01 to 2025-05-31), Hubble evaluated 104 valid candidates across three rounds with zero runtime crashes. The top five discovered factors spanned range, volatility, and price trend families, avoiding the common pitfall of converging on volume-only motifs. Out-of-sample validation on a held-out period (195 trading days from 2025-06-01 to 2026-03-13) showed that the two range factors and two volatility factors retained strong predictive performance with statistically significant HAC-adjusted t-statistics for Pearson IC and long-short returns, while the weakest trend factor decayed materially. The framework also demonstrated robustness across different LLM backends. Hubble provides a reproducible, safe, and interpretable workflow for alpha research, with persistent diagnostics artifacts enabling post-hoc analysis and financial interpretation.