Co-champion (tied with s11) and best validation loss of the study (2.628). b12c152nbt-pat (4.21M params, 2686 MFLOP/eval, ~30.8 single-thread CPU-ms): s11 resumed +100k steps on the full pool with an lr warmdown tail. Beats g170e-b10c128 by ~+5.6 +/- 2.8 judge scoreLead (pooled 192 games, 48 visits, b18 judge 256v); h2h vs s11 is +1.3 +/- 4.3 — i.e. the polish improved validation loss but was Elo-flat vs s11 (a clean val != Elo datapoint: the cheap continue-polish does not transfer once the architecture's capacity bottleneck is relieved). Full checkpoint (optimizer included, resume-capable).

Details + the full study writeup: https://github.com/sanderland/vibego (experiments/WRITEUP.md). Distilled from public kata1-b18c384nbt over katagoarchive.org positions — credit to lightvector and the KataGo distributed-training contributors.