CYBER-HOBO
concepts
in practice
how it is built

The human in the loop

In an AI-augmented practice, attention is the high currency. The model's output scales. The human's attention does not, and everything that matters passes through it: the human is the throat of the system. The value of the whole arrangement is decided there, so that interface is designed and measured like any other part of the system.

Where drift happens

Drift is not random. It has places. Late in a long session, errors compound quietly, and correcting one leaves in place the conditions that produced it: starting fresh with a clean state document works better than arguing with a loaded session. At the start of a session, corrections hold: calibration lands best at boot. The hardest case is mid-task, with everything loaded and the work moving. There, written rules do not interrupt the motion. What catches the moment is a control that forces a question at the point of action: a permission asked, a check refusing, a decision put before the act.

at the start WHAT HAPPENS drift is low, instructions fresh WHAT WORKS corrections land and hold mid-task, loaded WHAT HAPPENS silent choices inside the motion WHAT WORKS a forced question at the act late, long session WHAT HAPPENS errors compound quietly WHAT WORKS a reset, clean state document

Spending attention

The design consequence is that attention is budgeted. Review goes where a claim carries risk, as the trust page describes. Decisions are put as explicit options rather than left implicit in prose. Reading load is treated as a cost: work arrives as documents read at the reader's pace, and what a page asks of its reader is watched. These are cognitive choices.

Measuring the collaboration

The collaboration itself is measured. Three quantities carry the picture:

correctionshow often the human had to fix the assistant's course
unsolicited proposalshow often the assistant volunteered work nobody asked for
reading loadhow much text the human must read per session

The measurements exist because this matters. What augmented work does to the human, to attention, to load, to judgment over years, is an area with little research data, and this system is built to bring data to it. The research that exists is young: a 2026 study by Lepine and colleagues measured cognitive load in AI-assisted professional work in a controlled setting, and records like this practice's add the longitudinal, real-work side that controlled settings cannot see. These pages are part of that argument.