Designing Student Rockets for the Worst-Case 1% of Flight Scenarios: Checking an LLM-Agent-Built Pipeline Against Actual Flight Data, with Re-implementation by a Second LLM Agent

From 2022 to 2026, 187 of 524 rockets launched in the International Rocket Engineering Competition scored zero recovery points. Yet none of 15 student design documents showed every rule holding in all but the worst 1% of thousands of simulated flights with varied wind, thrust and mass, and the rulebooks reviewed ask only for fixed cases. Such verification needs a large software pipeline, so the author had the LLM agent Claude Code build one that accepts a design only if all implemented numerical flight and recovery rules (mostly the strictest limits of three rulebooks plus project rules) meet that standard in 9,600 flights in archived winds. Six of 13 missions passed, two with assumed stand-in motors, and on new random seeds 5 of the 6 kept a 95% lower bound of at least 99% on the share of flights passing all rules. To see whether this verification matters in design, the agent swapped the design search's Monte Carlo flights for 5 fixed-input flights, with equal margins and sizing. The bound then fell below 99% in all 9 runs, while the full method reached at least 99% in 6 of 9. Nearly all these failures came from one closely binding rule, the 2.0-caliber pad stability margin of NASA Student Launch. Because these bounds cover only sampling error and rest on the flight model, the weather and the Knacke/Pflanz loads, the agent compared the models with flight logs. With forecasts made 6 or more hours before launch, mean absolute peak-altitude error was 3.60% on 8 archived flights and 5.02% on 9 pre-registered ones, but the one measurable parachute opening shock was at least 2.9 times the prediction. To audit the code, Codex, a second agent, reproduced all 13 missions file by file and re-implemented the recovery sizing and descent statistics without reference code or results, matching all 222 values. No designs were built or flown, and the contribution is the LLM-built method and its audit.