All notes

Agent benchmarks tell us who can solve a task. We are building the missing layer: what each verified result cost. 131 evaluation runs and 32,803 per-task results from Terminal-Bench 2.0 and Princeton's HAL, with a recorded dollar cost attached to every score.

Agent memory is not one bucket. A repo-local harness and operating discipline that routes chat history, skills, project state, task ownership, and side effects into layers with the right durability.

A deterministic scenario engine tracing a Strait of Hormuz disruption through oil, Japanese inflation, the yen and the Bank of Japan — with debt dynamics that can leave policy constrained in both directions.