Changelog · Prove · 2026-10-29
Prove 0.1.0: the first release
0.1.0
Poisson Prove tells you whether a change to an AI agent really changed how often it fails. It
replays the same cases before and after the change, on the same seeds, compares the two runs
condition by condition, and gives a verdict: safer, worse, trade or no significant change,
with stated error bounds and detection limits. When no change counted and the runs could not have
shown one of the size you accept, it says too small to tell; a sweep planned that small is
refused, with the number of replicates that would do. Python 3.11 to 3.14, on macOS and Linux. It
is open source, under the MIT licence.
Start. pip install poisson-prove, then poisson-prove init --example. The example needs no
key: it sweeps a mock agent and a patched version of it on the same seeds, and
poisson-prove diff reads trade.
What’s in it. The poisson-prove command with ten verbs (init, sweep, diff, check,
report, paired, serve, inspect, runs, ui); check as a gate for CI; a terminal app,
opened by poisson-prove with no verb; a dashboard of static pages built from the store; and
init --agent, which writes a skill and a prompt for your own coding agent to set Prove up in
your repository for you to review. Sweeps run on your machine, in Modal containers on your own
account, or, experimentally, on your own Burla cluster in your own AWS, Google Cloud or Azure
account (the [burla] extra, pinned to Burla 1.8.2).
What it does not do. Prove makes no model call of its own and never phones home. The dashboard in 0.1.0 is static pages, rebuilt by the commands that write to the store; it does not update while you look at it.
Next. A live local dashboard over the same store, then setting up and running sweeps from either the terminal app or the dashboard.