Guard

Gate outputs, bound authority to verified competence, protect access, plan recovery, and defend human development.

The family calendar automation you built has quietly become load-bearing: it schedules the kids' activities, emails the babysitter, and last Tuesday it confidently sent everyone to a canceled swim practice across town. Nobody caught it because nobody checks it anymore. The system got authority faster than it earned trust.

GUARD is a five-part fortification you run on any consequential component, human or automated, before and during its operation. Gate outputs: test what it produces before acting on it, using samples or independent checks. Uncertainty and authority bounds: cap what it may do at its verified competence, captured in the governing rule that authority should not exceed verified competence and autonomy should not exceed recoverability. Access and provenance: give it minimum necessary data and keep records of sources and versions. Recovery and resilience: build the exception queue, rollback, and manual fallback, and actually rehearse the degraded mode. Development and dependence: protect your own ability to function without it. The cue is any component whose failure would propagate before a human noticed. Applied to the calendar: spot-check its output weekly, remove its ability to email third parties until accuracy is proven, keep a paper fallback. What changes is that failures become detectable and recoverable instead of silent. Skip heavy guarding only where mistakes are cheap and instantly visible.