empirical-validation

Identify what should be true after the change, run the appropriate command or action to observe it, capture the actual output, and confirm it matches the expected outcome. Record the evidence in the task journal rather than relying on code inspection.

Use a browser subagent to navigate to the affected page and capture a screenshot showing the expected visual state. Confirm specific elements like headers, loaded data, and layout correctness instead of assuming the component renders properly.

Valid evidence is actual captured output: API responses from curl commands, build success logs, screenshots of UI state, or query results showing expected data. Statements like "this should work" or "the code looks correct" are explicitly forbidden.

The task is not marked complete, the failure is documented in the state file, and a fix task is created if the cause is known. After three or more failures, a context health monitor is triggered to reassess the approach.

It should always block completion when verification is possible; the protocol has no bypass for verifiable changes. It integrates with verification and execution workflows to ensure every completed task has documented evidence.