RESEARCH & WRITING

Notes on building efficient models and agents.

The architecture, geometry, and math of what we build.

RESEARCH NOTE · ARCHITECTURE · JUNE 2026

The Geometric Rotation Transformer

How we replace a model's weight grids with rotations, and why that makes it more compact and adaptive, with structural stability that holds before a single step of training.

READ THE FULL NOTE →
RESEARCH NOTE · SYSTEMS · JUNE 2026

The orchestration layer above the loop

Claude Code and opencode are excellent agents — but they live inside one loop. Production needs what sits above it: durable pipelines, approval gates, checkpointing, and reliability metrics. We map the five canonical agent workflows onto a declarative engine that runs any CLI agent as a composable step.

READ THE FULL NOTE →