Research
Original research and technical investigations — cross-domain problems, rigorous methods, reproducible results.
4 papers
We compared void radius distributions from DESI observations (DESIVAST DR1, V² with VIDE pruning) against ΛCDM predictions from AbacusSummit N-body simulation. Three non-parametric tests agree: no significant difference. Clean null the standard model's predictions hold for VIDE-defined voids at low redshift.
Reconstruction benchmarks often assume that method performance is determined primarily by algorithm family. Using a cosmic matter density field from the AbacusSummit N-body simulation suite, we demonstrate that the same physical field and identical observations produce opposite winning methods depending only on representation. We formulate a two-channel recoverability model separating low-rank algebraic structure from spatial correlation structure, and confirm its falsifiable prediction: constraining the field into a low-rank regime reverses the winner.
Matrix completion — the math behind Netflix recommendations — fails 5x worse than standard interpolation on ocean temperature data. The reason is a one-paragraph proof about what the algorithm can't see.
A V2 re-audit of the original tornado-ERA5 draft found that the headline signal should not stand. Eleven rerun packets plus five verdict-only packets retired the original claims across five mechanisms: post-event leakage, multiplicity, catalog sensitivity, source absence, and confirmed negatives. This rewrite replaces the original narrative with what survived, what was retired, and how future AI-assisted research should be gated at n=0.