principle-prove-it-works

Verify task completion by checking the real artifact directly: build the code, run the actual feature path, and confirm data flows from input to output. Building successfully is necessary but not sufficient proof that the feature works.

Verify delegated work by inspecting the actual output artifact, such as the git diff, file contents, or runtime behavior, rather than the agent's summary. Agents report what they intended to do, not always what actually happened.

Indirect verification through file mtimes, output freshness, cached screenshots, or self-reports can reflect stale or derived state rather than reality. When verification fails, suspect the observation method before suspecting the system itself.

The strongest proof is a deterministic script that re-runs the same comparison, such as diffing old and new compiled output. Keep the script's output visible as an artifact a reviewer can re-run instead of trusting a one-time visual check.

Commit verification artifacts only for large or complex work where the trail must be auditable later, such as a big port or migration. Most routine work only needs the artifact visible to the human reviewer, not committed.