Written ≠ Wired: Hunting Fake-Green in Agent Systems

A green light is a claim, not a fact. This preprint names four species of fake-green — the defense that was written but never wired in, the process that exists but isn't running, the exit code 0 that reports done without doing, and the check that measured the wrong surface — and documents four measured incidents from a working AI-agent fleet, July–August 2026: a nightly off-machine backup that reported success for eighteen consecutive nights while its copy step failed every time; a pair of protections that failed open exactly when their precondition was missing; an operations note that claimed an alarm was dead while the scheduler said it was loaded; and a model-CLI adapter that returned exit 0 together with zero bytes of output. The doctrine is counterfactual verification: every protection ships with a test that removes or mutates it and watches the failure happen; verification reads the live surface, never the note; "could not measure" is a stop, not a pass. A protection you have never watched fail has never been tested. Mutation testing, chaos engineering and dead-code detection are mature disciplines; the narrow, newly urgent claim here is that these instincts must now be turned on generated defenses — the guards an agent writes and reports as done — because that is where fake-green is produced fastest. Note on authorship. The work was produced inside an AI-agent ecosystem and is about that ecosystem's own failures. The named authors of record are the humans who built it; the AI contribution — spine, body, and verification gates, plus an adversarial cross-family review panel — is disclosed openly in the article's author's note. Version 1.1 adds the Turkish edition (written-ne-wired-TR.pdf / .md) — a faithful adaptation, not a re-derivation: all measurements, code strings and citations are identical to the English original; the narrative is in Turkish while technical terms remain in English.