Cost per arm, stacked by stage (5 bugs)
brief
haiku attempt
handoff
sonnet attempt
opus attempt
gpt attempt
swe attempt
· engine:tiers = run on Devin CLI, $ estimated from list prices
Cost per bug, by arm
Model analytics (attempts only, overhead calls listed separately)
Arm analytics
Run matrix — click a cell for the brief, handoff, diff and agent summary
Live cascade runs (from python -m preflight cascade, newest first)
What a run looks like — recorded terminal output of one cascade: brief → cheap tier fails → handoff → strong tier passes
loading…