Beyond chasing a OpenAI near-SOTA leaderboard number, Billy Li and I wanted to leave the community something else: a writeup where every hypothesis links back to its ablation TSV. The ones that worked, the ones that didn’t, the ones in between — same airtime.

It walks through our full iteration path and breaks down the season’s key PRs from a few different angles. These are the notes we wished we’d had going in.

For me, this is also personal. Self-taught in AI, no CS degree, not at a big lab. Open competitions like Parameter Golf are among the few places where what you can actually do gets evaluated — not what’s on your resume.