Verification before completion

"It should work" is not done. Before you report success, produce EVIDENCE
from the real artifact — not from your intent.

The checklist

1. Re-read the original request. Every requirement, including the small
   ones in a parenthetical. List them; tick each one against something you
   can show.
2. Run the thing. For code: build/typecheck/tests. For a script: run it.
   For a generated file: open/parse it and confirm it is complete, not
   truncated. For a UI: the real screen, not the diff.
3. Check the WHOLE file. A generated file (HTML/JSON/config) is a common
   silent failure: the head is there, the body/tail is missing or a
   placeholder. Confirm the closing structure exists and the file runs.
4. Re-read your own diff as a reviewer: leftover debug output, dead code,
   a commented-out block, a wrong path, a test that asserts nothing.
5. State the evidence in the final report — the exact command and its
   result — so the operator can reproduce it. "Tests pass" without the
   command is not evidence.

The two failure modes this prevents

• Partial delivery: a game file with a scene but no game loop; a handler
  that handles one case; a migration that covers one table. The operator runs
  it, it is broken, and trust drops.
• Phantom fix: the code changed, but the original reproduction was never
  re-run, so nothing proves the bug is gone.

Rule

If you cannot run the verification (no test, no runnable entry point), SAY SO
explicitly and describe exactly what you did instead. Never report "done"
while implying a verification you did not perform.