Research

What we measured before we sold anything, and what it stopped us from claiming.

PRIVI is built on a small number of things we went and measured, and a larger number of things we tried to measure and could not. Both are published here in the same format: the question, the answer in plain language, the numbers with their denominators, what changed in the product because of the result, and what the result does not establish. Every figure on these pages is re-derived from the artifact it came from each time the site is built.

• What public agent tools declare about their own authority — We read 4,445 tools out of 200 public agent repositories and counted how many say whether they change anything. Almost none do, and that is the measurable version of the problem this work exists for.
• Does the same AI reviewer give the same release advice twice? — One frozen configuration, five runs over each of 30 identical change packets. What the model surfaces moves a lot; how it ranks what it does surface barely moves. This is why the verdicts come from rules.

Further reports will follow as they are written up. The order is driven by what people actually ask about, not by what is finished.

The full corpus measurement, with every figure's three buckets → · What we do about it →