A log, not a manifesto.

This site collects experiments on the scaffolding around language models — tool surfaces, context policy, loop control, guardrails and the evals that tell you whether any of it worked. Every claim here comes with a run behind it.

Posts are written as run logs rather than as advice: the setup, the single variable under test, the numbers it moved, and the traces that let you check the work yourself. Where a result did not hold up, that is written down too.