open measurement for agent skills
Batuta is an open measurement layer for Agent Skills: local routing, an explicit control group, and public results tied to the raw evidence.
Local benchmark recorded at audited baseline 9fa7471 on August 24, 2026. These are provenance-bound baseline measurements, not ecosystem results or current-device guarantees.
Batuta Zero
A dependency-free Rust binary indexes installed skills and ranks matches locally with BM25. The measurement layer is designed to work with other routers too. Silence is a valid result when the match is weak.
A declared holdout keeps the router silent on a configurable share of turns. Comparing outcomes with and without a suggestion is more useful than confusing correlation with lift.
Published rankings require multiple installations and carry counts, model context, protocol version, and stated sample bias. An empty ranking is more credible than a polished anecdote.
open source · reproducible
The repository contains the protocol, frozen battery, schemas, implementation, and the commands used to reproduce the measured baseline.
LAB · phase 0
Phase zero publishes bounded aggregates from trusted LAB runs. Prompts, secrets, signatures, receipts and project content never appear here.