amarbaroAMARBARO labs
Measured, local-first AI engineering,
published with receipts.
amarbaro.org is the research side of the AMARBARO group: inference engines, agent tooling and the measurements behind them. The commercial side lives at amarbaro.com.
Latest research
22 measured write-ups, each with a claim and the result that would falsify it. All research
Practice · 2026-09-23
A Reboot Is Not a Health Check
The production stack came back from a reboot with 37 of 37 checks passing only after the check covered containers, Cloudflare routes, and the browser…
Models and quality · 2026-09-23Krea 2 Wins the Background
On the same nursery prompt and three seeds, Krea 2 produced the more useful background composition in 2 of 3 comparisons, while Z-Image kept sharper…
Practice · 2026-09-23The Cache Is Not the Only Token Budget
Filtering redundant hook context before it reaches the model removed 964 bytes from a real payload, reducing it from 1790 to 826…
How we work
- Every claim names its falsifier. A result states what would prove it wrong, and the measurement that would show it.
- Receipts stay on disk. Numbers trace to a commit, a log or a result file that still exists. No receipt, no number.
- Local first. The work runs on hardware we own: one workstation, one GPU, no rented cluster behind the result.
- Failures are published too. Inconclusive and wrong predictions stay in the record; a list of only wins is marketing.
- A check that meets the thing. Done means a test that exercises the artifact the way a user meets it, not a green build.