NeuroGolf fourth-place solution: agent loops, shared idioms, and runtime budgets

In the fourth-place solution to Kaggle's NeuroGolf competition, a network with very few parameters could still be too slow to run. The team's writeup says some operators, notably Einsum, could use only a few parameters yet take hours to run all examples on a single CPU. Runtime therefore became a constraint separate from a model's score.

The team's agents each worked on one task at a time and repeatedly improved it, while a registry retained multiple versions of each task's model by hash, with their scores, local evaluation times, and leaderboard status (unknown, correct, or broken). For the final submission, the team approximately maximized the total score of the selected models under a budget defined by their summed evaluation time. Keeping alternatives mattered: the writeup says choosing every task's highest-scoring model could produce a submission too slow to finish. It is a concrete example of optimizing an entire workload rather than each component independently.