The dirty secret of AI-driven quantitative research is that the model writing the experiments can also poison the evidence. A leaky feature that scores well in a backtest gets stored as a successful precedent, then propagates through every later iteration. Prompt-level guardrails don’t fix this because the author and the reviewer share the same blind spots. Researchers from Princeton, Ant Group, and Stanford are proposing a structural fix: make the cheating impossible, not just discouraged.

Their system, AQuA, pairs two independent research loops that share no agents, no memory, and no candidate spaces. One discovers symbolic alpha factors on crypto markets; the other develops time-series models on US equities. The trick is what the team calls asymmetric freedom. The agent explores freely inside a constrained domain-specific language, but every split, feature definition, label, and evaluator is frozen before any iteration begins. The evaluator sits outside the adaptive surface. What improves across runs is the research process itself, not the scoring function.

Part I runs a six-agent pipeline for factor discovery on five-minute crypto data. Every handoff goes through an orchestrating Manager, keeping runs auditable. Factors enter as falsifiable proposals with hypotheses, mechanisms, and refutation conditions before any expression gets assembled. Across 20 research epochs, combined validation Spearman IC climbed to roughly 0.190, versus 0.171 for an adapted AlphaMemo and 0.151 for AlphaGen. Individual factors stay weak—single-factor ICs of 0.026 to 0.037—so the headline number is about the harness, not any one expression.

Part II predicts thirty-minute forward returns on intraday US equities using a config-diff approach: one hypothesis equals one architecture, loss, sampler, or optimizer change. The hybrid model—convolutional front-end, LSTM/Mamba/attention backbone, cross-sectional mixing—posts a per-stock raw IC of +0.0843, a 37.5% relative gain over the best baseline. Sector-neutralized and volatility-targeted, the held-out Sharpe reaches +2.50, with positive years from 2021 through 2025 including the 2022 drawdown. That consistency is the part worth paying attention to. Plenty of quant research produces a great backtest and a mediocre live year. A system that stays positive through a rough market suggests the integrity boundary is doing real work, not just looking good in a table.