hillclimbs

Configs, run receipts and leaderboards for public hill-climbing experiments (kernel autotuning, Blender reference matching, prompt/curriculum optimisation, harness speed). One prefix per climb: <climb>/config.yaml, <climb>/receipts/, <climb>/leaderboard.json. Every run here uses public data only (public evals and benchmarks); each receipt lists its input datasets. The same pipeline code also runs privately on personal data, and those results are never published here.