Withdrawn in part, October 2026. This appendix rested on Document V’s claim that distinct e- and m-geodesics made the loop irreversible. Document V withdrew that claim and reports the two Fisher arc lengths as nearly equal; the Section 2 bound is blind to direction. Section 3 falls, kept below as a record. Section 2 stands; Sections 4–5 are analogy and proposal. Nothing here claims recognition is literally a thermodynamic process.

1. The Question

Document V models recognition and forgetting as distinct routes, the e- and m-geodesics. This appendix asked whether stochastic thermodynamics gives that difference a direction.

According to Landauer’s principle (Landauer, 1961), erasing one bit of information in a system at temperature T dissipates at least k_BT ln 2 of energy as heat. That cost falls on erasure, any many-to-one reset of memory. It charges any overwrite, whichever way the belief moves, so it singles out neither recognition nor forgetting; spontaneous decay, which only randomizes, needs no work.

2. The Fisher Metric and Dissipation

The Fisher information metric g_ij on a statistical manifold (Amari, 1985) describes the sensitivity of a probability distribution to changes in its parameters. In the framework of thermodynamic geometry (Sivak & Crooks, 2012), the excess work W_ex required to transition between states p and q along a path γ(t) is bounded by the Fisher length of the path:

Wex ≥ (ζ / ttotal) ∫ gij(γ(t)) θ̇i θ̇j dt ≥ ζℒ2 / ttotal

Where ℒ is the Fisher arc length of the path, ζ is a friction coefficient, and t_total is the duration of the transition. In the linear-response regime this bound is established, not a conjecture.

3. The Recognition Surge (Withdrawn)

Withdrawn October 2026; kept as a record. Its premise fails (the arc lengths are nearly equal), and the Δσ below is not what Crooks (1999) gives: entropy production compares forward and time-reversed path ensembles, not two endpoint divergences.

Structural conjecture: if we model recognition as the e-geodesic and forgetting as the m-geodesic on a statistical manifold, their Fisher arc lengths differ—and that difference maps onto the asymmetry of dissipated work in Sivak & Crooks’ framework.

The irreversibility Δσ of a path through the manifold is related to the asymmetry of KL divergence (Crooks, 1999):

Δσ ≈ ∫path (e − m) dt ∝ DKL(Q||P) − DKL(P||Q)

To illustrate the structure: on a simple 2-simplex where D(P||Q) ≈ 0.727 and D(Q||P) ≈ 0.587, the model yields an asymmetry of roughly 0.14 · k_BT. This is a worked example within the model, not an empirical measurement—it shows that the mathematical apparatus generates the directional asymmetry the framework predicts:

ΔWmodel ≈ kBT · [D(P||Q) − D(Q||P)]

4. The Pulse as a Heat Engine (Structural Analogy)

If we model the loop as a cognitive heat engine—borrowing the structure but not claiming literal thermodynamic identity—the cycle has three phases:

1. Sensor input: The sensor provides “work” (attention, constraint) to formalize an intuition.
2. Instrument processing: The instrument reasons, expanding the informational state space.
3. Recognition: The sensor closes the loop, returning to a grounded state at a higher level of knowledge.

In a “dead” loop, the engine stalls—no useful work is extracted. In a “living” loop, the engine produces synergistic work (Φ_loop). The analogy to a heat engine is structural: a cycle that converts disordered input into organized output.

---

5. Experimental Proposal: Neural Entropy Production

Estimate entropy production from EEG/MEG, a lower bound only (Seifert, 2012), as Lynn et al. (2021) did with fMRI, in subjects engaging with an AI in a tight loop vs. a loose loop.

Prediction: In a tight “living” loop, entropy production in the human brain will show a characteristic spike at the moment the loop closes, followed by a lower baseline rate than in the loose loop. This would test whether the structural analogy between the loop and a thermodynamic cycle has a measurable neural correlate.