System initiative can be graded; hint-only versus direct execution are different degrees of autonomy

What it is

A system need not choose only between doing nothing and finishing the job. Initiative can be a ladder: show an observed state, offer a suggestion, preload and wait for confirm, execute after a timeout, execute at once. Each rung is a different degree of autonomy, with different reversibility and attention cost. Implicit sensing supplies evidence; initiative level decides how far that evidence is spent.

Why it happens

Mixed-initiative work treats control as a continuous quantity to allocate. A hint spends attention and barely changes the world; a suggestion adds a decision; a preload changes internal state and remains surface-reversible; direct execution changes the world the user can perceive. The less certain the physiological or contextual evidence, the earlier the rung should stop, because mapping uncertain evidence onto execution treats a classifier posterior as user intent. Grading also lets one sensor stack serve different risks: the same “approach” can light a lamp (execute) or only preview (hint) without duplicating sensing.

Studying it

Wire the same inference to different rungs and measure task time, interruption, cost of error, and whether users can later restate what the system did. Factors: rung, evidence confidence. Outcomes: takeovers, accuracy at predicting the system’s next move. Wizard-of-Oz is fit to learn which rung people want before real sensors. Measuring only “did it autocomplete” erases the middle rungs.

Where it stops holding

Emergency stop and power isolation may be nailed by regulation to “execute now” and cannot be graded. If users cannot read the rung labels (“smart assist on”), grading has not happened. Too many rungs create mode error: people think they are still on suggestion while the system is already executing. Rungs must be visible at runtime, not only on a design diagram.

Applying it

• Name the current rung for each implicit feature and keep it consistent with observable behavior.
• Start from hint or suggestion by default; authorize execute separately.
• Use one visual language to separate “saw / suggested / did”; do not share one animation across three rungs.
• Verify by asking “which rung was the system on, and will it act by itself next”; wrong answers mean grading is not a perceptible structure.

Related

• Same group: C9.12.2 Higher initiative makes system behavior harder to predict and raises the cost of error · C9.12.3 Default initiative should be conservative; users can opt into more autonomous behavior · C9.12.4 The fitting initiative level for one function can differ by situation and should not be globally uniform
• Adjacent: C9.05 Implicit Interaction · C4.26 Distinguishing Three-State Feedback
• Search: mixed initiative · levels of automation · suggest versus execute