Author Reports 79.90% Accuracy for Gear-Evolved DSH on Titanic
In a local experiment, the author reported that Codex + Astra achieved a best accuracy of 72.96% on Kaggle Titanic under a one-hour limit. In a separate configuration where XGBoost was explicitly suggested and the agent ran for roughly 90–100 minutes, the best result was 77.27%. The author then used Gear’s GEPA loop to evolve a DSH harness with Astra at maximum reasoning effort. The first evolution round produced a harness with 79.90% accuracy, the best of two runs, on 418 local test rows; candidate runs averaged about 13 minutes of agent execution, excluding time spent generating harness changes.
In this reported local experiment, the Gear-evolved DSH harness scored higher than the other two configurations. The result provides a concrete example of evaluating and evolving the workflow around an AI agent, while remaining specific to the author’s local test and not establishing broader performance.