Large language models (LLMs) can generate SQL and SQL‑Python workflows for investment research, yet ensuring point‑in‑time (PIT) validity remains difficult. Agents must respect disclosure cut‑offs and select the correct historical states and record versions throughout data retrieval and financial analysis. To address this, we introduce BuySideFlow, a benchmark built on a fixed snapshot of a licensed production financial database, comprising 408 practitioner‑derived Chinese tasks that span equities, funds, fixed income, and macroeconomics. Executable reference implementations are provided, along with a taxonomy of temporal errors that pairs result checks with expert audits of the executed workflows. Evaluation of nine agent configurations across all tasks yields result accuracies ranging from 16.2% to 36.8%. An audit of 900 runs on 100 shared tasks reveals that 193 of the 380 runs that passed result checks (50.8%) violated PIT constraints, demonstrating that output agreement alone does not guarantee temporal validity.
Review: BuySideFlow exposes a critical weakness of current LLM agents in handling time‑sensitive financial data, offering a systematic evaluation framework and error diagnostics that chart a clear path toward improving temporal compliance.