AutoResearch Tabular automates tabular ML discovery

This is the right kind of agentic ML: constrained, auditable, and optimized for search throughput rather than one-off leaderboard theatrics.

• –Expanding time windows are the credibility anchor here; they mirror production better than k-fold and make leakage much harder to sneak in.
• –Limiting edits to `features.py`, `model.py`, and `analysis.py` blocks the classic "agent improves its own grader" failure mode.
• –`LOG.md` and `LEARNING.md` give the loop persistent memory, which is the difference between compounding insight and repeating dead ends.
• –LightGBM defaults plus feature/tree caps are boring on purpose: the real objective is more honest experiments per day, not one giant run.
• –This is best read as signal discovery on operational tabular data, not benchmark chasing on a fixed dataset, so the hundreds of runs only matter if the target is real-world improvement.