use disclosure: experiments were largely implemented and run by LLM agents; we directed the strategy, audited the results, and approved every submission.
What we submitted
A randomized QMC estimator in the Sobol family with antithetic sampling, plus standard implementation-level cost optimizations. Design choices were fixed by paired A/B tests under production randomization.
The accounting that held up
We worked with the identity raw x F = V_iid * c / g (per-sample IID variance scale, billed cost per effective sample, variance gain vs same-N IID). Calibrating V_iid ~ 0.0405 on public MLPs predicts the pure-sampling plateau at 6.24e-7; the observed value is 6.47e-7, so the model is off by 3.7%. Our production point is consistent with it too: 40960 samples x 1.91 measured gain / ~62.6k full-budget samples = 1.25, which matches the platform ratio. This is the same plateau arithmetic from the floor-bet thread (which the organizers later confirmed); what we can add is the measured calibration and the consistency check.
Negative results
Two things we tried that failed. Both turned out to be independent confirmations of results other teams published first.
- Control variates on top of Sobol. A cross-fitted linear CV worth 1.42x under IID keeps only 1.04x under our Sobol — both eat the same low-order variance. Matches hyojun_kwon’s “input-space variance reduction does not survive fusion” observation, found independently.
- Billing arbitrage: none found. We audited the evaluation-version cost model directly. All matmul/einsum/tensordot routes bill at ratio 1.000, fp16 = fp32, fp64 = 2x, and comparisons/where/take/concatenate are all priced. Whatever the frontier is doing, it isn’t metering tricks.
Summary
For judging: the calibration and consistency check of the plateau identity, and the two negative confirmations above. Ledger excerpts and probe scripts available on request.
Thanks to the teams whose write-ups made the frontier legible — several of our negatives are independent confirmations of things they published first.