Submission ID: 327651
Public score: 1.23e-7 adjusted / 1.55e-7 final-layer MSE
I am sharing the attached technical write-up for my selected Phase 1 submission.
The estimator combines a near-complete BCH strength-4 orthogonal-array design with a structure-aware forward pass. Positive homogeneity removes the Gaussian radial integral; a small pilot classifies neurons as dead, on, or kink; dead paths are pruned, late on paths are folded linearly, and uncertain paths remain fully sampled. This dead/on/kink execution chassis was inspired by the SOX team’s public write-up. The BCH first-layer product and one antithetic continuation step are then evaluated with exact Walsh–Hadamard identities.
The main statistical improvement came from the BCH/OA angular design. The final leaderboard gain came from exact FWHT, antithetic, zero-skipping, and fast-matrix-multiplication identities that reduced the arithmetic actually performed. Every numerical operation remained inside the official flopscope.numpy interface; no work was moved outside the meter.
The PDF also includes the main ablations, limitations, and negative results.
LLMs were used extensively for ideation, implementation, analysis, code review, and
drafting. I directed the experiments, reviewed the results and made the final decisions.
Attachment: phase1_writeup_submission_327651.pdf (66.5 KB)