AgentSeeResearch Notebook
version 1.0.0 · created 2026-04-08 · updated 2026-04-08

Actuator Categories

protocolhypothesisoriginal
ClaimThe AI's understanding capabilities ARE actuators. The actuator set includes communication modulation, observation surfacing, withholding, controllability provision, environmental signals, and silence.
This claim fails if
If actuator categories beyond this set are necessary for capacity maintenance, or if these categories prove insufficient to produce measurable state changes.

Key insight

The AI's understanding capabilities ARE actuators. How the machine communicates, when it surfaces observations, what it presents, what it withholds based on state -- these are all actuation modulated by both understanding and state estimation.

Categories (non-exhaustive)

  1. Communication modulation. Timing, framing, complexity, emotional tone of what the machine says.
  2. Observation surfacing. "What you're doing doesn't match what you said matters" -- timed to when the person can process it.
  3. Withholding. Deciding NOT to surface information because state data says the person can't process it. Sometimes the most important action is inaction.
  4. Controllability provision. Structured experiences where the human detects action-consequence contingency. This is the mechanism by which the vmPFC-DRN pathway is activated (Maier & Seligman 2016).
  5. Environmental signals. Breathing cues, biofeedback, ambient modification.
  6. Silence. Sometimes the most important actuator is doing nothing.

Steering tension

The question of what the machine DOES when detecting divergence during degraded state is where the backfire prediction and caring orientation intersect. Making departures visible remains an observer function ("what you're doing doesn't match what you said matters" is observation; "stop that, do this" is control). However, a machine that selectively times observations using its model of the human's values, with predictable behavioral effects, is doing something more complex than passive observation. The Socratic method is formally questions. Nudge theory formally preserves choice. Neither is accurately described as pure observation.

Whether the observation/control distinction can be maintained in practice -- or whether it collapses into sophisticated persuasion -- is an empirical question the architecture must answer, not assume. Kill conditions 6.4 and 6.6 test for this failure.

Systematization needed

A formal mapping from estimated state variables to available actuator inputs to predicted effects on state is required. This is the control design problem and depends on the plant model (see open-problem: actuator-systematization.md).