About BMS

Less theater. More evidence.

BMS is an independent publication from Kyokasuigetsu about what frontier coding agents can actually do.

Why BMS exists

Agent demos are abundant. Comparable evidence is not. We built BMS to make serious evaluation readable: rigorous enough for builders, clear enough for anyone deciding which systems to trust.

The name is deliberately plain. BMS is a bench, not a stage—a place to put systems under load, measure what happens, and keep the failed parts on the table.

What we publish

Our benchmarks focus on consequential engineering work: security review, repository-scale implementation, debugging, and other tasks where correctness matters. The journal covers the messy work around those tests—protocol design, grading, agent economics, and surprises that do not fit in a chart.

Independence and disclosure

BMS is produced by Kyokasuigetsu. We do not accept model-vendor control over methods, grading, or conclusions. If a benchmark is sponsored or credits are provided, that relationship will be disclosed on the publication itself.

Questions, corrections, and reproducibility requests are welcome. Public contact and repository links will be added when the publication repository is created.