Plan
R1 reads the task, surveys the codebase, and writes a human-readable plan before any write happens. That plan is the first useful review boundary, because it exposes scope and assumptions while the task is still cheap to redirect.
Execute
Execution is instrumented work, not a free-form loop. Every read, write, and retry is part of the receipt chain. The work can fan out, but it does not disappear into undocumented agent state.
Verify
This is where R1 earns the name. Tests, lint, policies, and hostile review happen here. If the run cannot defend its own result, it does not proceed.
Commit
The merge boundary is serialized and conditioned on the gates. A run that passed earlier still has to survive the final check after rebasing and before the receipt is sealed.