All notes

An explainable EIA-based oil-inventory stress index ran from 41 in April to a peak of 96/100 the week of Jul 3, then eased to 93 as products rebuilt while crude kept drawing. Every point is attributed to a specific inventory series.

Agent benchmarks tell us who can solve a task. We are building the missing layer: what each verified result cost. 131 evaluation runs from HAL, Terminal-Bench, and curated observations; 32,803 raw Terminal-Bench task results collected.

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.