Research hub
Metal kernel optimization
The tracking and publishing surface for the program. Weekly updates, curated reports, and one page per completed work item.
Start here
Weekly Update
The current readout for anyone who only has time for one page.
The newest update gets summarised in this slot and archived as a dated page under weekly/. The steps for publishing one, including the files that have to be touched, are in README.md at the repository root.
Curated evidence
Reports
Standalone writeups that carry a finding worth reading on its own.
Qwen3.5-35B-A3B Metal-kernel capability evaluation
Final report of the frozen qwen35-metal-eval-v1 campaign. A 35.9 percent one-shot visible-correct rate, a 3.236x morton speedup that four sessions reproduced but the stability gate refused to confirm, and the three fixes required before any RL pilot.
Benchmark walkthrough
How both evaluation tracks work and exactly how to run them. The 12-task Metal-Sci benchmark with paired ABBA timing and hash-chained records, plus the 215-task Gimlet-style KernelBench-to-Metal reconstruction, the hosting setup, and the gate order.
Work record
Runs
One page per completed work item, appended as items finish.
Completed work items get their own page under runs/ and a card in this section. New cards are spliced in at the insert marker above, and this block is deleted by hand in the same edit that adds the first one. See README.md for the procedure and runs/ for the full index.
Orientation
What lives here
This repository is the writing surface. The engineering code and the bulk data live elsewhere.
Writeups and tracking
Weekly updates, curated reports, and per-work-item pages, published as a static site. A push to main is the deploy. No training code and no kernels live here.
The kernels themselves
Implementation, benchmarks, and tooling live in the separate metal-sci-kernels repository. Pages here cite it rather than copy it.
Artifacts and dumps
Checkpoints, raw generations, and full metric dumps live in Cloudflare R2 under the bucket metal-kernel-optimization. Small per-item local evidence goes in artifacts/, which is gitignored and never deployed.
How this page is maintained. Each section above carries an insert marker, weekly:insert, reports:insert, and runs:insert, so new cards can be spliced in at a fixed point without reflowing the file. Cards use one status vocabulary, pass, watch, limit, and info, with the badge text written in uppercase in the markup. Adding content means touching this page, the section index, and, for a new newest weekly update, the alias in _redirects. Every section here is empty today, so the first entry added to a section also removes that section's empty state.