Andrew Fribush

fig. 2.15.1: sixty gate-passing cards, thirty per author, judged by the gate, the blind judge and you, author revealed after

fig. 2.15.1 · sixty gate-passing cards, thirty per author, judged by the gate, the blind judge and you, author revealed after.
QUALITY CHECK: 3% STUDENT VS 90% FRONTIER; GATE 97.5%ONE BAR IS THE SHARE OF AN AUTHOR’S CARDS A JUDGE PASSED, ACCENT FOR THEFINE-TUNED STUDENT; PASS OR FAIL EACH CARD TO JUDGE THEM YOURSELF.CARD 1 OF 60 · TESTED 电影 · FILM电影在晚上。The movie is in the evening.PASS IF IT IS GRAMMATICAL, MEANS SOMETHING, AND THE ENGLISH MATCHES.PASSFAIL60 CARDS, 0 READ; ACCENT IS THE STUDENT0%50%100%THE MECHANICAL GATEDETERMINISTIC JS, ONE SHOT, N=238 · 22097.5% · FINE-TUNED 0.6B56.8% · FRONTIER, FIRST ATTEMPTTHE BLIND JUDGEOPEN-WEIGHT 35B, 30 CARDS EACH, BLIND3%90%YOUNO CARDS READ YET0 OF 00 OF 0

The sixty cards are the ones the blind judge read: thirty from the frontier ladder’s bank and thirty the fine-tuned Qwen3-0.6B wrote at 475 iterations that passed the mechanical gate, sampled with the judge’s own seed. Every card here passed the gate. Your rubric is the judge’s (grammar, sense, translation); its fourth criterion, whether exactly one option fits, needs the options, which the card does not show. Who wrote a card is revealed only after your verdict.

source: minimodels DIARY.md:2267-2290 (blind judge, 30 vs 30), projects/cardgen/README.md:37-50 (gate rates), projects/cardgen/judge.py:96-110 (sampling, seed 0) · cardReader.data.ts

View in context: How a student model gamed its verifier → · All figures

Andrew Fribush, software and computer engineer in Austin. I build consequential systems where software meets institutions: finance, healthcare, commerce, government data, and the mechanisms that govern AI agents. Taking new engagements: hello@andrewfribush.com.