← Home

Research

My paper is “Characterize Before You Loop: What Transfers From a Zero-Shot Proposer’s Template and What Does Not.” When a language model proposes candidates inside an evolutionary search loop, does its preferred score still predict its behavior once the parent, the iteration, the serving path, or the vendor changes? I tested this on circle packing in the unit square, scored by an exact local evaluator. Short answer: not reliably. It is self-published and not peer reviewed.

By the numbers

  • 8 registered arms
  • 1,255 planned slots audited; 293 valid at a 1e-9 tolerance, and none beat the family maximum by more than 1e-6
  • 0 of 105 valid packings from bare pinned-API requests, which shows how much serving configuration matters
  • Anchor matches fell from 11 of 17 valid seed outputs to 11 of 49 valid descendants across 4 loop lineages
  • 26 of 26 valid outputs met keep-or-improve under a better in-family parent
  • A 20-lineage, 600-slot replication on a second model
  • 15 pages plus a 22-page supplement, with code, ledgers, and preregistrations public

Links