verify-with-controls
Installation
SKILL.md
Verify With Controls
A reading is not a finding. Before "no matches" (a zero), "half the checks vanished" (an alarm), "the work landed" (a success claim), or "nothing is running" (stillness) becomes a claim, the reading must survive its controls. Those are evidence attached, a positive control behind every zero, the instrument suspected, the postcondition checked, a re-measure at the moment of sending, and a checker that has been seen failing.
Six rules, each distilled from a false conclusion nearly shipped in production:
| # | Rule | The failure it kills |
|---|---|---|
| 1 | Evidence or error, never a bare zero | one 0 for both "absent" and "I failed to look" |
| 2 | Positive control before trusting a zero | a blind instrument answering the wrong question truthfully |
| 3 | Suspect the instrument first | a plausible number measuring the wrong thing |
| 4 | The postcondition is the truth | success logs over a state that never changed |
| 5 | Two observations to look, re-measure to send | a report racing the state it describes |
| 6 | Negative control for checkers | a gate that cannot go red |
1. Evidence or error, never a bare zero
A probe that can return 0 both for "absent" and for "I failed to look" is broken by construction. No amount of care in reading its output fixes it. The two meanings arrive as the same byte.