CrashGym is a deterministic, replayable limit‑order‑book simulator. An LLM agent must choose between a deleveraging intervention and a liquidity backstop for a market that is about to crash due to a shock. Before committing, the agent can query intervention experiments in a laboratory version of the market. The final score is computed on an independently seeded target realization where the simulator executes every feasible intervention, allowing regret relative to the best intervention to be measured directly without human or model judging. The simulator implements four amplification channels—persistent price impact, liquidity withdrawal, margin‑driven liquidation, and an evolving strategy mix—each of which can be toggled on in stages. In a prospectively specified study across 24 newly sampled worlds, the only deployment with a complete record (Luna) showed mean regret dropping from 8.878 under observe‑only to 3.785 with laboratory access, a reduction of 5.092 utility units (95% world‑bootstrap interval 1.028–9.507), i.e., 57.4% of the observe‑only mean. Prospective comparisons for two additional deployments are still pending; an exploratory study found lower mean regret with laboratory access for all four evaluated deployments. A laboratory whose strategy mix evolves endogenously showed no clear advantage over one with a fixed mix. CrashGym therefore measures the value of experimentation as avoided decision loss and lets the simulator’s detail be judged by the decisions it supports.
Review