Finance agents ground events as low as 0.8 out of 5

Four leading LLMs scored near 4 out of 5 for plausible investment logic but only 0.8 to 2.8 out of 5 for grounding that logic in market events, according to InvestLogicBench’s 201,247 documented decisions from 151 investors. The mismatch means a profitable backtest can conceal reasoning that was lucky, temporally wrong, or inconsistent with the investor’s profile. Teams applying the enterprise AI investment thesis to financial agents should version profiles and evidence, keep decision ledgers, and score process separately from returns.