One LLM wrote a trading feature. Another reviewed it. Both missed a future-data bug

An AQuA research loop let one LLM write an intraday trading feature and a second LLM review it for causality. The feature divided the volume seen so far by the day's final total volume. At prediction time, that final total included trades from later in the day: future data. The reviewer approved the causal-sounding explanation. The feature then lost its unusual signal on a clean re-split, and a manual audit found the leak. That leaves a real choice: build a smarter reviewer, or give the agent a smaller, safer toolset? This is not simply “LLMs make mistakes.” Two model roles converged on a plausible explanation while the actual bug sat in the operation the authoring agent was allowed to express. A more capable causality review would still have to reconstruct the data flow behind that denominator. AQuA v2 responds structurally. The paper says the data splits, feature and label definitions, and evaluator are sealed outside the agent's adaptive surface. Arbitrary feature code is replaced with a fixed registry of causal operators. In that specification space, the full-day normalizer is not an available operation. That guarantee is bounded by the implemented registry and sandbox. It also creates a cost. A registry can make a known-invalid transformation impossible, but every restriction narrows what the agent can invent. A better reviewer preserves more of the search space, but has to catch subtle dependencies reliably. The paper is “AQuA: Recursively Self-Improving Quantitative Trading Research Agents” by Guo et al. (arXiv:2608.12841). Appendix B does not publish a numeric IC for the failed feature. What makes it worth reading is the documented path from a convincing false signal to an interface redesign, not a claim about live trading performance. If you were designing this loop, where would you spend the complexity budget: a stronger reviewer with static data-flow checks, a smaller typed toolset, or a hybrid? What failure would make you switch sides?