fig. 2.7.1: share of each pipeline stage’s cards the Sonnet re-screen rejected, by generator fig. 2.7.1 · share of each pipeline stage’s cards the Sonnet re-screen rejected, by generator.⛶ THE RE-SCREEN REJECTED 57% OF HAIKU CARDS, 12% OF SONNET CARDS ONE BAR IS ONE STAGE’S CLEAN CARDS, 0 TO 100%; THE ACCENT IS THE SHARE THE SONNET RE-SCREEN REJECTED, THE COUNT AT ITS END. 0% 25% 50% 75% 100% HAIKU-GENERATED · 217 OF 378 · 57% REJECTED FIRST DRAFT 126 OF 222 · 57% RETRY ON THE FIRST DRAFT’S FEEDBACK 91 OF 156 · 58% SONNET-GENERATED · 46 OF 377 · 12% REJECTED ESCALATION 46 OF 377 · 12% EVERY ONE OF THE 755 HAD PASSED THE HAIKU JUDGE; 263 FLIPPED (34.8%). ON THOSE 263 IT HAD SCORED NATURALNESS 5/5 ON 108 AND 4/5 ON 153, AND RECORDED ZERO FINDINGS ON 136.
source: vocabinator research/2026-08-24-cardbank-rescreen-root-cause.md lines 3, 9-12, 16 (repo e77a474), re-derived from data/cardbank-zh.db stage_history and data/cardbank-rescreen.json · judgeFlipRates.data.ts
View in context: Calibrating an LLM judge → · 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 .