Introduction
Proof learns the customer outcomes your product must protect, exercises them in a real browser, and turns successful AI exploration into deterministic Playwright coverage.
What Proof does
Proof starts from a business task, not a selector list. It observes the page, chooses browser actions from the current state, verifies the resulting customer outcome independently, and compiles the verified semantic trace into a Playwright test that runs without AI.
When that deterministic test later fails, Proof retries the original business intent. If the outcome still works, it can prepare a validated test repair. If the outcome no longer works, it reports a regression with the run history and captured evidence.
Ways to use Proof
The mental model
- Business flow: a customer outcome worth protecting, such as completing checkout.
- Scenario: one happy path or edge case with explicit success evidence.
- Exploration: an AI agent reasons through the scenario in a browser.
- Semantic trace: the verified, selector-independent record of what worked.
- Compiled test: a versioned Playwright test replayed without model calls.
- Adjudication: a fresh semantic attempt that decides whether a failed test is stale or the product is broken.
Execution starts paused
A new Proof environment will not create a Sandbox or call a model until its independent live-execution and AI interlocks are explicitly enabled. See Cost controls.