Prove it works

A result is checked by looking at the real thing it produced, never at a proxy for it. A wrong inference acted on costs far more than the look that would have caught it.

When: the changed path's result is read through a proxy (a timestamp, a status file, a cached value, a screenshot, or a report from another agent), a check fails, or the work is a port or a migration whose trail is audited later.

• A self-report, a subagent's summary, and "it compiles" are not evidence, per the default stance.
• Look at the thing itself: a running process is checked directly, not through state derived from it, such as a file timestamp, an output that looks fresh, or an old screenshot.
• Read the value itself, never a cached or derived copy of it.
• When a check fails, suspect the way it observed before suspecting the system.
• A comparison that reruns beats a single look. The rerunnable form is the project's test runner, build, or linter, and its saved output in the evidence record is the trail a reviewer can rerun.
• Show that saved output to the user, and commit it only when the work is large enough that the trail must be audited later, such as a port or a migration.

Where it stops: the proof is the project's own test runner, build, and linter, plus the changed path driven end to end; a script written on the same branch as the change never proves the claim, whatever it prints. A claim no command can settle goes to rubric.