verify

Evidence-based verification is enforced by a multi-step process requiring agents to run commands, read output, and independently confirm findings before making any completion claim. This prevents rationalization and ensures all actions are checked against fresh evidence.

The systematic checklist for validating code review findings is a 5-step process: Identify the claim, Run the relevant command, Read the output, Verify the evidence independently, and then make the Claim. This ensures all review findings are rigorously checked.

To verify that test commands validate passing tests, re-run the relevant test commands and confirm that the command output explicitly validates the claim. This mandates reading the output thoroughly to ensure the successful exit code matches the expected test results.

No specific testing frameworks or dependencies are required. The verification process operates by enforcing a rigorous workflow for executing commands and reading output, meaning it works with any command-line tools or testing commands already present in your environment.

You should use a mandatory verification process whenever an agent makes completion claims, review findings, or processed reports. It is specifically needed to prevent the propagation of unverified claims and maintain integrity before actions like merging code changes.

An agent needs to independently confirm evidence to prevent rationalization and ensure accuracy. Trusting command output without thorough reading and independent confirmation allows errors to propagate, whereas verifying fresh evidence maintains integrity.