Replays
Every run leaves a map.
Four real runs, seen from above. Higher ground is a better model, and the amber trail is the path of ideas each one kept.
Recorded without an LLM: ideas come from a built-in search.
01 / 04
Housing
19 ideas tried, 2 kept, stopped on its own.
1,000,342RMSE on data it never sawPredicts “price” (a number) from 12 columns of 545 rows.
02 / 04
Breast cancer
37 ideas tried, 4 kept, stopped on its own.
0.9936ROC-AUC on data it never sawPredicts “malignant” (yes or no) from 30 columns of 569 rows.
03 / 04
Iris with missing
21 ideas tried, 3 kept, stopped on its own.
0.1067Log-loss on data it never sawPredicts “variety” (one of 3 kinds) from 4 columns of 150 rows.
04 / 04
Wine
15 ideas tried, 2 kept, stopped on its own.
0.0230Log-loss on data it never sawPredicts “cultivar” (one of 3 kinds) from 13 columns of 178 rows.