The Coordination Persistence Conjecture

The Trust Attractor draws on dissipative thermodynamics, information theory, and models of coordination. The Coordination Persistence Conjecture, first drafted as a theorem, asks whether these strands compose into a universal persistence advantage for invitation. The derivation is incomplete. Its gaps include the relation between information and dissipation, the relation between dissipation and lifetime, preservation under coarse-graining, and the dynamics that would turn a higher accessible dissipation ceiling into an attractor.

What the Argument Can and Cannot Establish

Invitation preserves participants’ ability to choose. That gives a reason to value it, and a possible mechanism for adaptation when circumstances change. A universal thermodynamic ranking requires more: comparable joint options, actual transition rates, and a survival objective.

Present existence supplies no future survival objective. A system can exist now while approaching certain termination, or trade longevity for reproduction or another end. Even a valid survival theorem would recommend its strategy only to someone seeking that survival outcome under the theorem’s conditions. Hume’s is-ought gap remains beside the empirical question of which arrangements last.

The Thermodynamic Tools

Far-from-equilibrium driving can produce ordered instabilities under suitable dynamics, geometry, and boundary conditions. Heated fluid forms convection rolls only beyond its onset threshold. A spatially homogeneous driven reaction cycle can instead sustain steady concentrations and positive entropy production. Flow alone guarantees neither pattern formation nor a universal preference for coordination.

Landauer’s principle supplies a precise erasure bound. Erasing a single bit warms the world by a definite minimum amount, and by more in a hot room than a cold one. Precisely: an isothermal reset of an unbiased bit at bath temperature $T_b$ requires heat of at least $k_B T_b\ln2$ under its standard assumptions. For biased information the entropy being erased matters. Processing that erases nothing (logically reversible processing) is not taxed this way, so the bound sets no toll on every bit processed or observed.

Earth’s magnetic field is a useful case. The budgets of the geodynamo that generates it illustrate another necessary distinction: energy available to run a process and entropy production permitted by its gradients are different accounts. Comparing biological or social coordination requires explicit definitions and common units for both.

The Links Still Needed

Information and dissipation. Mutual information is a stock of dependence. Signed information flows and entropy production rates are different quantities. Positive dependence between driven subsystems need not create an additional dissipating current. Thermodynamic uncertainty relations constrain minimum costs of specified currents under their own dynamics; they supply no extra accessible ceiling simply because coordination is more precise.

Dissipation and lifetime. A fluctuation theorem compares a path with its own reverse. It says nothing about which of two unrelated arrangements lasts longer: rates 100 and 10 have the same forward-to-reverse ratio as rates 0.01 and 0.001, while reverse waiting times differ ten-thousandfold. Longer persistence needs its own dynamical argument.

Options and adaptation. Preserving every participant’s marginal options does not order joint options. Two agents can each retain two choices while only the pairs (0,0) and (1,1) are available. A different arrangement can restrict one participant and expand the other’s alternatives enough to have a larger joint set. Nor does bounded state entropy bound sustained entropy export: a driven cycle repeatedly traverses the same few states.

Scale. Coarse-graining discards detail. A lower bound on each system’s entropy production need not preserve their ranking. Two true rates 10 and 9 can have coarse bounds 1 and 8. Exact preservation in a specified model is valuable; carrying it across chemistry, biology, and society requires checking the same hypotheses at each level. Commutativity of combining and coarse-graining is algebraic compatibility, without a proof that a particular coordination advantage survives.

Dynamics. A high ceiling on accessible dissipation neither makes high-dissipation states reachable nor gives them significant probability. Transition rates, selection or adaptation rules, and an attractor definition are needed before differential persistence can be deduced. The published tools support a research programme; their assembly here supplies no universal invitation-dominance theorem.

The companion annex gives the conditional comparisons that remain sound and the missing inputs for a stronger result.

A parallel chain aims at a nearby destination through a neighboring formalization. A chain I will call the Amari Chain, drawn from Zhuravlev (2026, arXiv:2603.09067), starts from the same body of work the Trust Attractor draws on in Chapter 17, the programme that reads physical dynamics as learning dynamics in a neural network (Vanchurin 2022; Chapter 15 develops it). Its starting axiom is different: causal invariance. In Wolfram’s hypergraph physics, causal invariance is the requirement that the order in which the universe’s update rules are applied does not change the resulting web of cause and effect.

Apply the rules in one sequence or another, and the same events end up with the same causal relationships between them. (Imagine baking a loaf: whether you add the salt before the water or after, the finished bread is the same, because the steps that matter commute.) The Amari Chain traces a different route:

causal invariance → persistent observer (where a stable observer exists) → internal model (with a modeling mechanism) → Fisher information metric (for a regular statistical model) → natural gradient descent (for a chosen objective and Fisher steepest descent).

The persistent observer is the chain’s least self-explanatory link. It is any structure that holds together across the updates, keeping enough of itself intact from one rewrite to the next to have a stable vantage point on everything else. A whirlpool qualifies and a splash does not. Only something that lasts long enough to be updated repeatedly can accumulate a model of what keeps happening to it.

Responsiveness alone does not make a persistent observer an optimal regulator. The Good Regulator Theorem (Conant and Ashby, 1970) requires a specified regulation problem and its model and optimality assumptions. A Fisher metric additionally requires a regular, identifiable parameterized probability family. Neither the family nor the control objective follows simply from persistence. A thermostat and an immune system motivate investigating such models; each needs its own regulated variables and disturbance law.

For a differentiable loss and positive-definite Fisher metric $F$, natural gradient descent uses the direction $-F^{-1}\nabla L$. It is locally steepest under a Fisher-metric step constraint. This is a metric-dependent direction, with a step size still to choose; geodesic updates and other intrinsic methods also respect the geometry. Singular metrics need separate treatment. The mathematician Shun’ichi Amari, who gives the chain its name, supplied this local geometry in 1998; his result gives neither a unique globally optimal learning rule nor a direct rule for weighting evidence strength.

Useful internal models can improve energy capture when their predictive benefit exceeds the cost of learning and using them. A hunter that knows where the prey will be can spend less searching, and the difference is paid in calories. Thermodynamic competition can favor that net advantage; model accuracy alone does not determine it.

The two proposed chains ask complementary questions about coordination and learning. Neither establishes a universal physical selection law for invitation or an optimal update rule for every persistent observer. The convergence is suggestive, though its evidential weight should be discounted. Both chains draw on the same learning-dynamics literature, the line running from Vanchurin’s neural-network physics through its evolutionary and information-geometric offshoots (Chapters 7 and 15), so this is agreement between neighbors rather than independent confirmation. If neighboring formalizations do yield the same organizational pressure, that pressure is more plausibly a genuine feature of the landscape than an artifact of one derivation’s assumptions; a derivation from a genuinely unrelated tradition would carry far more weight, and finding one remains open work.

For the proposed synthesis, conditional results, and remaining proof obligations, see the online companion at https://www.thedeeperlaw.com/companion/annex/coordination-persistence-theorem/.

Notes

Notes for this chapter are available in the online companion at https://www.thedeeperlaw.com/companion/.