This is a working draft shared ahead of publication. It may change, and it carries no publication date until it is final.
Here's the problem. "Human in the loop" has become the seatbelt sticker of AI products — claimed everywhere, structural almost nowhere. In many systems the "loop" is a confirmation dialog bolted onto an architecture that would work identically without it. Under deadline pressure, bolt-ons get bypassed.
The checkbox version
You can spot the checkbox version by asking one question: if the human stopped participating, would the system still function end-to-end? If yes, the human is decoration. The system was designed autonomous-first, with oversight sprinkled on top for the enterprise slide.
The architectural version
In the architectural version, human judgment is a component other parts depend on:
- Approval gates are routing, not UI. Consequential actions — sending, committing, spending — physically route through a gate. There is no code path around it.
- Memory is reviewable because of how it's stored. If beliefs live in inspectable, attributable entries, review isn't a feature; it's a property.
- Tool use is observable by default. What ran, when, against which sources — recorded as the action happens, not reconstructed afterward.
- Handoffs are the unit of delegation. Agent work ends in a reviewable result, so review isn't an interruption of the workflow — it is the last step of the workflow.
IEVA's prototype learned this the honest way: as a personal system where memory was Markdown and history was Git, oversight was physically unavoidable. Every belief was a file you could read; every change was a commit you could inspect. The product direction treats that as the specification: scale the capability without breaking the property.
Why this survives contact with reality
Architectural oversight has a practical advantage over checkbox oversight: it degrades safely. On a rushed day, a checkbox gets clicked through. A routing gate still shows you the recipient and the commitment you're about to make. The system's guarantees don't depend on the human being at their best — which is precisely when guarantees matter.
Human-in-the-loop is part of the architecture. If it isn't, it's marketing.