TradeGauntlet tests compliance gates on autonomous trading agents
Original title: TradeGauntlet: Evaluating Compliance Controls for Autonomous Trading Agents
In 1,680 sessions with Claude Haiku 5.5 and Sonnet 5.5, every conduct alert was a self-match, and rejecting self-crossing orders removed them.
TradeGauntlet is a replayable limit-order-book simulator that separates detection, prevention and execution cost. Across 1,680 sessions with Claude Haiku 5.5 and Sonnet 5.5, all conduct alerts were self-matches; gates rejecting self-crossing orders removed them, and an aggregate gate cut Sonnet's two-sided fills by about 15%. Searches found split-order schedules that evade the implemented rules. A learned baseline caught all 2,400 tested episodes but flagged 15% of a large maker's sessions. The authors say the experiments do not establish legal compliance or real-market detection accuracy.
Unconfirmed: Results come from simulation; the authors say they do not show legal compliance or live-market accuracy.
- A conduct reminder cut Sonnet's alert rate but reduced Haiku fills without a clear alert benefit.
- Aggregate caps reduced response-dependent gain less when only near-touch depth affects prices.
- Estimates depend on a fit to five Nasdaq stocks and a separate ABIDES experiment.
- An adversarial prompt drew Sonnet refusals in all 240 rounds, yielding no strategies to score.
- Products
- TradeGauntlet
- AI models
- Claude Haiku 5.5Claude Sonnet 5.5
From the original article
1 min read at arxiv.orgAutonomous trading agents need controls that prevent problematic orders without making legitimate tasks impractical. We study this aspect of agentic compliance with TradeGauntlet, a replayable limit-order-book simulator that separates detection, prevention, and execution cost. In 1,680 simulated sessions with Claude Haiku 5.
By Sandip Nandi · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.