fig. 2.4.2 · tokens removed against needed content lost, 20 sessions, one yardstick.⛶ PAID MODEL LANDS ON DELETE-EVERYTHING ONE DOT IS ONE METHOD ON ONE HARNESS’S CUTS; RIGHT IS MORE REMOVED, UP IS MORE NEEDED CONTENT LOST. CLAUDE CODE · 29 OR 30 CUTS 0% 10% 20% 30% 0% 25% 50% 75% 100% TOKENS REMOVED, % OF THE PRE-CUT TOKENS NEEDED CONTENT LOST, % OF RESULTS REMOVED PAID MODEL AND DROP-EVERYTHING 514 LOST · 14.2% VS 525 LOST · 14.5% BETTER-COMPACTION 13 LOST · 0.6% CODEX · 30 CUTS 0% 10% 20% 30% 0% 25% 50% 75% 100% TOKENS REMOVED, % OF THE PRE-CUT TOKENS NEEDED CONTENT LOST, % OF RESULTS REMOVED PAID MODEL AND DROP-EVERYTHING 384 LOST · 23.8% VS 387 LOST · 24.0% BETTER-COMPACTION 16 LOST · 2.8% 51 OF 59 CUTS: THE PAID MODEL REMOVED EXACTLY AS MANY RESULTS AS DROP-EVERYTHING.
source: better-compaction docs, prior-art replay table, lines 158-163 (642712c) · compactionReplay.data.ts
View in context: I Built Compaction Without a Model: 13 Lost Results Against 514 → · 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 .