This is my Phase 2 algorithmic contribution write-up. The full 9-page PDF is attached below.
Submission ID: 328274. That is the single successfully evaluated Phase 2 submission this write-up refers to.
What it is
A fixed 0.75 / 0.25 blend of two estimator branches: gain-covariance propagation, and whitened antithetic Monte Carlo. On the full public Mini cohort it takes adjusted final-layer MSE from 4.050381542e-07 down to 3.044010045e-07, a 24.85% reduction, at 9.345% mean compute utilisation with 0 of 100 failures.
Why it works, and what I can actually claim
The measured branch-error correlation is r = 0.0382. Two nearly orthogonal residual fields give a quadratic benefit when combined, and that part is validated. The closed-form optimum weight is 0.7341 against the 0.75 I preregistered and shipped, a 0.118% excess over optimum that I deliberately did not retune after the fact.
The deeper story, that one branch mainly carries Gaussian-closure bias while the other mainly carries sampling noise, is plausible and consistent with the design, but no experiment isolates it causally. The write-up labels that an unvalidated explanation rather than quietly promoting it to a result. Section 1 defines the measured / derived / unvalidated taxonomy used throughout.
Negative results are in there too
A Hermite correction came out 1.2% worse than baseline. There were also operational runner and seed failures that cost real runs, and those are documented rather than dropped, because they explain the shape of the search.
LLM use
Disclosed in section 9, and not softened. The development process was LLM-driven under a registered experiment ledger: mechanism selection, code, hypotheses and falsifiers, preflights, failure diagnosis, residual analysis, and this write-up. Numerical claims were checked against finished structured records, file hashes and run identifiers rather than accepted from free-form model output. No access to the private evaluation suite, and no result inferred from the leaderboard.
Happy to answer questions on the blend identity or the weight-response measurement.
whestbench-phase2-algorithmic-contribution-328274.pdf (87.7 KB)