Andrew Fribush

fig. 2.15.2: every system as gate-clean yield against blind-audited quality, with usable per contract as isolines

fig. 2.15.2 · every system as gate-clean yield against blind-audited quality, with usable per contract as isolines.
MASK + VERIFIER, NO TRAINING: 82.5% USABLE; FINE-TUNE 3%A DOT IS A SYSTEM, LABELLED % USABLE; ACCENT IS THE PIPELINE, STEPS 1-5.QUALITY, % OF GATE-CLEAN CARDS A BLIND AUDIT PASSEDYIELD, % OF CONTRACTS WITH A GATE-CLEAN CARD50%60%70%80%90%100%60%70%80%90%100%0%60% USABLE75% USABLE90% USABLEFINE-TUNED 0.6B · 97.5% YIELD, 3% QUALITY, ~3% USABLEFRONTIER ×1 · 54.5%FRONTIER ×3 · 94%1 MASK + CODE, K=4 · 44.8%2 FEW-SHOT, REPAIR · 60%3 CONTRACT LADDER · 80%4 STRICT VERIFIER · 75%5 VOTES, PLANNED RETRIES · 82.5%

The number beside each dot is usable cards per contract, yield times quality. The fine-tune: blind 35B audit of 30 gate-passing cards. Other systems: blind cross-model audit of every gate-clean card in the deck, source hidden, read by the project’s Claude agent, a different model family from the 35B that generates; the frontier’s 96% is its 28 gold cards in the first of those audits. Step 5 is 35 of 40 contracts and 33 of 35 cards, so 33 of 40 is the 82.5%. The quality axis is broken below 60%.

source: minimodels projects/cardgen/README.md:37-50, 140-142, 216, 366, 417; DIARY.md:2276-2290, 4077-4078, 4127-4129, 4159 · yieldQuality.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.