Immutable
release. Only release title and notes can be modified.
Bounded calibration-pilot release for transparent local-first coding-agent routing.
On six calibration tasks, local-first reduced observed hosted inference cost by 45.56% versus the paired Sonnet 4.6 cloud-only condition, while resolving 2/6 tasks versus 3/6. This triggered a no-go for the original confirmatory design.
This release contains calibration evidence, not a population estimate or a general claim that local-first routing is inferior. It preserves valid and invalidated evidence, frozen treatment metadata, analysis, the pilot manuscript, and reproducibility tooling.
Archive SHA-256: 8279f0f396568c007a027395cfe81fe459b57a2188d608cff0425cd837773058