An autoresearch-style loop (modify, verify, keep or discard) that tunes a recommender on its own. Task: predict each user's next Mart visit (bodega, deli, pharmacy, convenience store) from their Food, Mart, and Shops history, using the public Foursquare NYC check-ins (1,080 users, 4,124 venues). Each iteration changes one thing; a paired bootstrap keeps it only if the gain is not noise; the winner is scored once on held-out visits. Code and live runner: GitHub.
This page shows the final run: run 5 of 5 (heuristic proposer, resumed from run 4's best, 50 iterations; logs of all five runs are in runs/ on GitHub). To run the loop yourself: uv run experimenter.py or uv run --extra app app.py for the Gradio app.