Auto-research with codex: How I achieved a 232x Faster Kernel

qr/ ├── submission.py # the live entry ├── submission_*.py (560) # named submission variants (crystal_rain, blue_reply, …) ├── modal_b200_*.py (119) # Modal B200 probe / compare scripts │ ├── AGENTS.md # top-level notes / logs ├── attempts_log.md ├── claude_ideas.md ├── leaps.md ├── problem_statement.md │ ├── docs/ ( 68) # per-experiment writeups & status docs ├── code/ # the actual QR kernel source tree ├── scripts/ # summarizers, timing, submit helpers ├── archive/ # cleaned-out old submissions & probes │ ├── submissions_20260616_17/ │ ├── submissions_20260618_20/ │ ├── probe_scripts_cleanup_20260627/ │ └── … ├── submit_logs/ # logs provided by the evaluator for each submission └── profile.*/ # captured NCU profile runs