Declaring the agent layer (who owns $Q$)

• Specify the agent/system, its model $Q$, sensors, actuators, goals, and constraints.

Closed loop and counterfactual sensitivity

• Demonstrate interventions that change future Texts.
• Design contrasts to reveal if the model’s predictions match realized data.

Model improvement metrics

• Predictive log‑loss on held‑out or next‑step data.
• KL divergence $D_{KL}(P \|\| Q_t)$ trends (with $P$ the source and $Q_t$ the agent model); negative free energy reductions.

Cost accounting

• Energy, time, bandwidth, memory; include in objective functions.
• Report Pareto fronts and robustness under budget changes.

Reporting standards

• Baselines: no‑feedback, random policies, fixed policies.
• Ablations: remove components (e.g., phase control) to quantify contribution.
• Misses: where the dual view failed or added nothing; refine the agent model first.

Further reading: Agency & Delegation, Research Directions, Glossary.