Case study: Two LLMs missed a future-data bug in coding and review loop

niacolhealth · reddit · 2026-08-25

A paper on AQuA (quantitative trading research agents) reveals a typical AI system vulnerability, highlighting blind spots in dual-LLM collaboration.

The Bug: One agent wrote a feature dividing volume-so-far by the day's final total volume. The denominator included future trades (data leakage). A second reviewer agent approved the causal-sounding explanation. The feature failed on a clean re-split, found only via manual audit.
Root Cause: It wasn't just a simple mistake; two model roles converged on a plausible explanation while the actual bug sat in the operation the authoring agent was allowed to express.
AQuA v2 Fix: The system structurally responds by sealing data splits and feature definitions outside the agent's adaptive surface. Arbitrary code is replaced with a fixed registry of causal operators, making invalid transformations like the full-day normalizer impossible at the interface level.
The Trade-off: A registry prevents invalid transformations but narrows the search space. A better reviewer preserves space but must catch subtle dependencies. The author asks: where should the complexity budget go—stronger static review, smaller toolsets, or a hybrid?

More from coding & agent