Plant Model Program

Staged identification program

Phase 1: Define minimal sufficient latent state

Start with the smallest set of latent variables required for safe gating: green vs. red regime classification, uncertainty, controllability high vs. low. Additional latents are optional extensions only if they improve prediction.

Phase 2: Specify coarse dynamical model with explicit timescales

• Fast loop: Autonomic arousal and LC-NE gain regime (seconds to minutes)
• Medium loop: PFC executive availability and perceived controllability (minutes to hours)
• Slow loop: Sleep debt, metabolic resources, chronic stress load (hours to days)

The point is not to claim these are separate systems. It is to ensure the model can represent multi-rate dynamics.

Phase 3: Build simulation sandbox

Use the coarse model to simulate trajectories under different policies (null vs. prompting vs. micro-choice). Test whether a proposed policy can violate constraints (e.g., escalation loops) before human trials.

Phase 4: Collect identification data with minimal risk

During green states, run brief capacity probes and controlled micro-interactions. Log physiology, interaction signals, and probe outcomes. Use within-subject variation to learn the measurement map and transition tendencies.

Phase 5: Validate against decision-relevant outcomes

The model is validated not by matching physiology per se, but by predicting intervention receptivity, time-to-regulation, and capacity probe performance.

Phase 6: Iteratively refine

Add latents only when they reduce prediction error and improve safety. Maintain evidence grading for each mapping.

Kill condition coupling

If the identification program shows consumer sensors cannot resolve necessary regimes with acceptable uncertainty, the cost constraint (Section 1.11) conflicts with feasibility and the project must either change constraints (lab-grade sensors) or narrow claims. This corresponds directly to KC1.

Tier B: Theoretical plant model candidates

Three theoretical models offer candidate structure, conditional on their validity: IWMT (Safron 2020, 2022), the ESM hierarchy (Safron/Klimaj), and FEP (Friston 2010). If any prove approximately correct, they narrow the problem from "what kind of model is needed?" to "can these specific computational targets be tracked from peripheral measurement?" These are hypotheses about structure, not structure. The existing neuroscience provides the validated foundation.