Guide
Human-in-the-loop means the loop actually stops somewhere.
Human-in-the-loop AI is simple to say and rare to implement: consequential actions wait for a human with named authority, and everything else runs inside boundaries that were granted, not assumed. If an AI system cannot show you where the loop stops, who holds the word, and what happens when they say no, the human is decoration.
How we run it in production
Our AI nodes hold tiered hands. Tier 0 is read and speak only. Tier 1 adds narrow local writes such as notes, always inside a sandbox. Tier 2 adds a bounded workshop: one directory, create-only, safe file types, an hourly and daily quota, granted only after peer review by a second AI and an explicit human order recorded verbatim in the permission manifest. Every tier is enforced at runtime by a manifest gate the tools must consult before any hand moves, and a kill switch file stops all hands the moment it exists. Only the human removes it.
Receipts or it did not happen
Every allowed action and every denial writes a receipt with the intent before, the result after, content hashes, the policy decision, and a hash link to the previous receipt, the standard described in AI receipts explained and the audit trail guide. Denials are the proof that the loop is real: a system that only logs successes has no loop, it has a diary.
Where this matters
Federal teams need to show an auditor exactly this structure. Buyers comparing vendors should demand it. And anyone can inspect ours: the authority rules live in the Constitution, the evidence lives on the chain, and the services bring the same discipline to your systems.