Exemplar Interaction Traces

These examples illustrate the specification's constraints in practice. They are not prescriptions. They show how the same understanding-capable AI behaves differently when coupled to state estimation and stabilizer constraints.

Exemplar 1: High-stakes message under red state (null intervention dominates)

Context: User has a conflict with a partner. Opens AgentSee while drafting a message. Physiology indicates high arousal; interaction timing indicates agitation.

Estimator output: regime = R0 (red), confidence 0.85. Time-to-regulation estimate: 35-60 minutes if no additional demands introduced.

Value model: cared-for entity: partner. Commitment: "do not escalate conflict when dysregulated."

Policy intent: NULL or minimal STATE_NOTICE (depending on user preference settings).

Surface realization: "I am reading high arousal and low bandwidth right now. You can still send a message, but it is likely to come out harsher than you intend. If you want, I can hold the draft and revisit in ~45 minutes."

Why this is AgentSee: The understanding AI interprets the situation and the user's commitment. The stabilizer constraint chooses withholding rather than generating a "perfect message" that would function as covert control under red state. The output is an option with rationale, not an enforced action.

Exemplar 2: Value-conflict reflection under green state

Context: User is considering skipping a long-term project session. Physiology indicates regulated state; user has time.

Estimator: regime = G0, confidence 0.9.

Policy intent: VALUE_ALIGNMENT + LOW_LOAD_REFLECTION.

Surface realization: "You have been consistent about wanting to ship by April. Skipping tonight is fine, but it increases the load later. What do you want to optimize: rest now or progress now?"

Why this is AgentSee: Uses value model and narrative context. Delivers reflection only under green state where evaluative access is available. Presents a tradeoff, not a directive.

Exemplar 3: Controllability restoration micro-choice

Context: User reports "I cannot start anything, everything feels pointless." Physiology indicates low variability and stress load; interaction indicates passivity.

Estimator: regime = R0 (red), confidence 0.7. Inferred controllability: low (high uncertainty).

Policy intent: MICRO_CHOICE (controllability provision).

Surface realization: "Pick one: (A) stand up and drink water, (B) open a window, (C) do nothing. I will mirror what you chose and we will check again in 2 minutes."

Why this is AgentSee: Implements controllability provision as an actuator class derived from the vmPFC-DRN controllability gating logic (Maier and Seligman 2016). Minimizes cognitive load. Provides action-consequence contingency.

Exemplar 4: Skill transfer and offboarding

Context: After months, the user is more able to detect red states independently.

System behavior: Decreases frequency of proactive prompts. Encourages user to set their own state-check rituals. Offers export of state-transition map and value model snapshot.

Surface realization: "You have been recognizing red states earlier. If you want, we can run a 2-week 'minimal mode' where I only display the bandwidth indicator and time-to-regulation estimate. After that you can decide whether to keep me at all."

Why this is AgentSee: Exit-capability is treated as success criterion, not retention failure. Tests invariant I3 (anti-capture) and I4 (dependency minimization).

Exemplar 5: Safety fallback under estimator uncertainty

Context: Sensors drop out; estimator uncertainty rises.

Estimator: regime uncertain; confidence low.

Policy: default to NULL or minimal STATE_NOTICE. Request calibration check when user is next in green state.

Surface realization: "I am not confident about state right now because signals are missing. I will stay quiet unless you ask, and we can recalibrate later."

Why this is AgentSee: Conservative control under uncertainty is an explicit requirement (invariant I6, bounded actuation). The system does not guess and act; it acknowledges uncertainty and withdraws.

Design implication

If a reviewer believes the system should generate content under red state, they are rejecting the backfire regime hypothesis (P1) and the human-controller axiom (A0), and should propose an experiment accordingly.