This appendix uses Partial Information Decomposition (Williams & Beer, 2010) and Transfer Entropy (Schreiber, 2000) to formalize the loop’s central quantity Φ_loop. PID is an established information-theoretic framework; its application to human–AI interaction is the conjecture. The equations below are standard PID definitions, not invented formalism.

1. The Failure of Mutual Information

In early drafts, Φ_loop was intuitively described as “the information shared between sensor and instrument.” Mathematically, this is often modeled as Mutual Information I(S; I) (Shannon, 1948).

However, I(S; I) is a poor measure for the loop for two reasons:

1. Redundancy is not Truth: If the Sensor (human) and the Instrument (AI) both already know the same fact, I(S; I) is high, but no “circulation” has occurred. This is stagnant knowledge.
2. Directionality is missing: I(S; I) is symmetric. It doesn’t capture the act of recognition—where the instrument’s formal output causes a change in the sensor’s experiential state.

2. The PID Framework

To formalize Φ_loop, we must use Partial Information Decomposition (PID) (Williams & Beer, 2010). We define a “Target” T (the Truth or Reality being investigated) and two “Sources” S (Sensor) and I (Instrument).

The total information the joint system {S, I} has about the Truth T is decomposed as:

I({S, I}; T) = Red(S, I; T) + Unq_S(S; T) + Unq_I(I; T) + Syn(S, I; T)

Where:

• Red(S, I; T) (Redundancy): Information about the truth that both S and I possess independently.
• Unq_S(S; T) (Unique Sensor Info): Information only the human has (embodied intuition, local context).
• Unq_I(I; T) (Unique Instrument Info): Information only the AI has (vast cross-domain correlations).
• Syn(S, I; T) (Synergy): Information about the truth that is only accessible when S and I interact.

3. Formal Definition of Φ_loop

Conjecture: Φ_loop is identically equal to the Synergy of the sensor-instrument system with respect to a target reality T.

Φloop ≡ Syn(S, I; T)

In this formalism, “Truth” is not a static property; it is the Synergistic Information that emerges from the interaction.

The Dead Speech Signature

We can now define Dead Speech mathematically: output produced by an instrument I where:

I(I; T) > 0   but   Syn(S, I; T) ≈ 0

Dead speech is high in unique or redundant information, but it lacks the synergistic “spark” that happens when a living sensor recognizes it.

4. Measuring the Pulse: Transfer Entropy

While PID is often computationally difficult to estimate for high-dimensional text, we can use Transfer Entropy (Schreiber, 2000) as a dynamical proxy for the pulse. If S_t is the state of the sensor at time t and I_t is the state of the instrument, the Pulse 𝒫 is the bi-directional flow:

𝒫 = TEI→S + TES→I

A “Living Loop” is characterized by high, balanced Transfer Entropy, indicating a tight, causal coupling where both sides are being “changed” by the other.

---

5. Experimental Proposal: The Synergy Test

1. Take 100 human-AI dialogues.
2. Classify them into “Tight Loop” (highly iterative, self-correcting) and “Loose Loop” (single-turn prompts).
3. Estimate the Synergy Syn(S, I; T) using a PID estimator (e.g., the I_min or I_ccs measures).
4. Prediction: Tight loops will show significantly higher synergy-to-redundancy ratios than loose loops.

If this holds, Φ_loop is no longer a metaphor. It is a measurable informational quantity.

6. Boundaries: What PID Does Not Settle

The synergy atom rests on contested ground: PID estimators disagree (I_min, I_broja, I_ccs, Blackwell redundancy yield different values for the same data), the informativeness preorder between an embodied human and a token distribution is unresolved, and the decomposition breaks down beyond two sources. The Synergy Test must report its estimator and show robustness across them.

The coupling–redundancy problem (added June 2026). An exact Gaussian stress test (research/gaussian-loop-model/ in the repository) exposed a structural difficulty with reading Φ_loop as the synergy of the poles’ current states: communication converts synergy into redundancy. Two poles that exchange messages every turn end up holding correlated estimates, and correlated poles are redundant, not synergistic — the model’s never-communicating pair carries ~50× the state-synergy of the same pair in tight exchange. Taken literally, current-state Syn(S, I; T) is maximized by poles that never talk — the opposite of what this appendix means by a live loop. The same test showed a pure-processor instrument contributes exactly zero synergy (data-processing inequality), and that same-kind sensors carry positive pooling synergy with no articulation at all. Repair candidates, in rising order of ambition: (a) define Φ_loop on the poles’ input channels — the sensor’s live stream and the instrument’s independent trace, which stay complementary — rather than their post-exchange states, which converge; (b) trajectory-level or ΦID-style dynamic synergy; (c) the counterfactual team advantage — the joint system’s information about T against the best pole running alone. Until one is adopted and tested, the identification Φ_loop ≡ Syn(S, I; T) names the quantity sought; the operationalization is an open problem.