DocsExamples
Examples
Each example was run as written, against the published tool in a fresh virtual environment, and every output and figure on its page is what that run printed and wrote.
- Run Prove on your machine Two sweeps of the keyless mock agent on one machine, then the comparison. No key, no account, no network.
- Run Prove on Modal The same two sweeps in parallel containers on your own Modal account, about a cent each, the cost printed first.
- Check a change on runs you already have A store that already holds two runs. List them, compare them, gate on one event, and the same from Python.
- Generate tasks, then check them with Prove Tasks' scripted Prove route: a pre-registration, three sweeps of the reference policy, the comparison. No model called.
- Gate a pull request with Prove A gate that passes, one that fails with exit 4, and the workflow that runs it on every pull request.
- Run Prove on your own agent A rollout around a keyless stand-in for your agent, a prompt change swept on the same seeds, and the comparison.
- Check a model edit on evals you already ran Stored eval results for a model and two weight edits of it. Replay them, pair them, see which edit broke tool use.
- Find what a model edit changed, and what it left alone An edit meant to stop over-refusal. See which prompts it fixed, and how big a benchmark fall it could be hiding.
- See how far you can quantize a model before its tool calling breaks One open model at full precision, 4-bit and 3-bit, run live on 30 tool tasks. 4-bit holds. 3-bit stops passing.
- Where does my controller break? A flight controller swept through wind gusts from every side, and the gust speed where half its flights crash.
- Can a fintech back-office agent move to a cheaper model? A payments operations agent on Claude Sonnet 5.5 against Haiku 5.5: where the cheaper model holds and where it does not.
- Catch a planted regression in a fintech back-office agent A payments operations agent on Claude with one tool broken on purpose: Prove flags it, Trace names the step it broke at.
- Check whether the same request gets the same decision A decision model sent the identical request twice: count the decisions that flipped, from stored responses.
- Check whether two leaderboard neighbors actually differ Two neighboring leaderboard entries compared on the tasks they share: one pair separates, one does not.
No example matches these filters.