Break is available: it finds where your agent fails before your users do. Read the launch
Prove is available: it tells you whether a change to your agent changed how often it fails. Read the launch
Poisson Labs

Changelog · Prove · 2026-11-04

Prove 0.1.3: a run in which every episode errored is named, and what to inspect first

0.1.3

A sweep in which every rollout errored suggests inspect with --limit 3. The sweep’s next: line does, so the first look shows three of the failing episodes rather than every one.

diff, report and check name a run in which every episode errored. They say so in plain words and suggest inspect of its commonest error with --limit 3, one line per such run, in place of an error about the event set. report writes nothing for it.

report ends with a next: line, poisson-prove ui. It is on the store that was typed, so the chain of next: lines has no dead end; --json carries the same command.

runs, the dashboard and the terminal app show such a run as stopped. A complete run in which every episode errored is shown as stopped, as the sweep did. Its manifest and runs --json keep complete, with rows and no_outcome carrying the fact; a partial run stays partial. The app’s comparison screen takes complete runs, so it no longer offers such a run.

CI runs Poisson Tasks’ own Prove check against Prove built from the tree. Before a release, and on every pull request, it runs the check from Tasks’ latest release, unedited, on a wheel built under a version the package index cannot hold, so the check cannot quietly test the release already out. A red check stops the release.

MATH ASINFRASTRUCTURE.

Talk to us