Measurement Stack

Layer P: Physiological Proxies (observer inputs)

HRV, EDA, pupillometry (if available), motion/activity feed the state estimator. These are inputs to inference, not outcome measures of agency. Outcome validity is assessed by whether these signals improve prediction of capacity-relevant outcomes in Layer C.

Layer C: Capacity Probes (objective-ish)

Brief executive function tasks provide repeated measures of whether the user is inside the envelope:

• Working memory probe (short n-back variant)
• Inhibition probe (go/no-go)
• Cognitive flexibility probe (task switch)

Constraints: must be brief (seconds to 2 minutes), must avoid becoming a stressor, must be scheduled only under green or green-fragile states.

These probes are not the definition of agency. They are measurable indicators of state.

Layer E: Evaluative Access Self-Report (subjective, structured)

Because agency includes access to "authentic desire," subjective report is required but constrained:

• Rate "can I think clearly" and "can I access what I want" separately
• Collect post-hoc ratings after the state passes (reduce red-state confounds)
• Treat red-state self-reports as lower-trust labels

The lower-trust status of red-state self-reports has a specific mechanism: interoceptive processing accuracy -- the anterior insula's capacity to translate bodily signals into conscious state awareness -- degrades under the same stress conditions that trigger the red regime (Wang et al. 2019). The measurement stack is therefore measuring a system whose self-report channel is impaired by the condition being measured.

Layer A: Action Coherence (behavioral)

Evaluate whether improved regulation and envelope time translate into:

• Reduced frequency of post-hoc regret reports
• Improved stability of commitments endorsed under green states
• Reduced impulsive pattern episodes (as defined by the user, not by external normative targets)

Wanting/liking divergence (Berridge & Robinson 2016) provides a specific pattern for Layer A to track: behavioral persistence toward rewards that no longer produce satisfaction or that the person has endorsed wanting to stop.

Validation

Together these four layers triangulate. No single layer is sufficient. The stack as a whole gets validated against the plant model program.