Integrated formulation and concrete models

The four substantive layers remain admissibility, efficiency comparison, weight and process. The work blocks below add their context, mutual consistency and test protocol; they do not introduce six alternative fundamental laws.

#A. Context and effect

Fix the system boundary, observables, initial conditions, resources, effect criterion, duration and tolerance. Distinguish the physically allowed space Ω_C from the subset Ω_E,C that carries the effect.

A state space may not be called an “arbitrary set” and at the same time be given, without further structure, a tangent bundle, derivatives and Euler–Lagrange equations. Discrete operations, smooth trajectories and stochastic paths require their own formal objects. A diffusion path is usually not differentiable; the old notation ρ(γ,γ̇,t) is not directly usable there.

#B. Operation and progression

Describe what can happen now, what actually happens and what context it leaves behind. For this, EEML-W provides the distinctions CAN, DOES and LEAVES. A seed is preparation; a cycle restores a named context; residual demand counts additional support during execution.

For a sufficiently completely described Markov system, a transition kernel K or generator L can specify the progression. The initial distribution and the environment are also needed. If the chosen visible state has memory, an ordinary time-homogeneous Markov generator on that state does not suffice: extend the state or use explicit history dependence.

On a graph, a cycle is only a possible continuation. Reachability, positive transition rates, stability and any exit routes determine whether it actually keeps running. The older tropical cycle calculations remain models of repeated execution at specified costs and duration; they do not yield a constant of nature.

#C. Input, support and durability

Keep a cost vector as long as there is no substantiated conversion between kinds of input. Compare at equal effect and with the full relevant context. A cheap intermediate step that makes the next operation impossible must not win against a slightly more expensive preparation on the basis of its low local cost.

Where appropriate, a path cost can be built up from:

This is an accounting possibility, not a requirement that every kind of effort adds up locally. Hidden memory must have been made explicit, and contributions must not be counted twice. Where needed, final reserves, stored damage and future possibilities also belong in the comparison criterion.

#D. Weight

Where the positive statistical description is justified:

The measure, S, ε, normalization and the regime of application are explicit. When macro-classes are compared, the macro-measure ν is added, as worked out in Part IV. This does not yet give a free model parameter a physically derived probability distribution.

Under independent composition, both costs and measure must have the required factorization. Under dependence, the couplings remain. Positive weighting does not replace coherent quantum amplitudes; Born is not derived from the multiplicative Cauchy rule.

#E. Coherence between the components

Weight and process are tested jointly. If π is meant to be the stationary distribution of the chosen dynamics, then L*π = 0. That is a consistency condition, not a unique determination of L. Detailed balance is a stronger, non-universal requirement.

Where a stationary small-noise limit exists, the exponent in the specified weights must moreover match the global rate function of the dynamics, with the same measure and noise convention. This is the delimited Freidlin–Wentzell bridge. Equality of the quasipotential with S does not prove reversibility: Part V gives an exact rotating Gaussian process with the same S and nonzero cycle current. Hard physical constraints are retained; a chosen goal need not automatically be reached permanently.

Exact reduction must yield not only the same static weights but also the same relevant visible trajectory probabilities. Write P for the full trajectory law and O for the observation map. Then, for every relevant visible event A, the following must hold:

The resulting visible process can have memory. Static log-sum-exp regrouping is one particular component of this broader gluing rule. The existing relation between reduction, memory and fluctuations has been worked out, for example, by Mori (1965).

#F. Reconstruction and testing

Measure or specify several aspects of the same process: occupation, directed transitions, waiting times, response and resource use. Look for a jointly fitting model and keep alternatives that cannot yet be distinguished. Then test a different load, blockage or initial condition without recalibrating.

Keep minimal closing sets of conditions, negative controls, errors and undecided outcomes. Self-application means that EEML's cost choices and explanatory freedom are also tested in this way. It is not yet a proven unique metamathematical fixed point.

#Genuine minimal resistance, with fixed effect

The old water/current intuition has a precise realization. For two fixed linear resistors R₁,R₂ > 0 and fixed total current I, it follows from current conservation and equal voltage that:

For a current-conserving deviation δ, the power difference is:

Here the minimum actually follows from the local laws and their coherence. The current uses both routes. This is Thomson's well-known principle in a simple case; see Doyle and Snell, §1.3.5.

The limit of this statement becomes apparent with nonlinear response: for different nonlinear responses, the stationary network minimizes a potential derived from the response, not necessarily the heating power. Precisely these results, both positive and negative, make it a useful benchmark.

#First concrete retrodiction: water through two parallel tubes

Known mechanical benchmark, added on 28 September 2026. Two rigid round tubes connect the same inlet and outlet, at equal height. Their inner radius r is the same; tube A has length L and tube B length 2L. We assume stationary, fully developed laminar flow of the same Newtonian fluid, no slip at the wall and negligible losses in the connections. The geometry is measured independently.

The well-known Hagen–Poiseuille law gives \[ R_i=\frac{8\mu L_i}{\pi r^4},\qquad Q_i=\frac{\Delta p}{R_i}. \] Therefore \(R_B=2R_A\). The common pressure drop and flow conservation fix the distribution: \[ \boxed{Q_A=\frac23Q_{\rm tot},\qquad Q_B=\frac13Q_{\rm tot}.} \] For a total throughput of 30 ml/min chosen in advance, we therefore predict 20 ml/min through A and 10 ml/min through B. That is the concrete comparison with an existing law of nature, not a new measurement. The local flow law and its conditions of validity are physical input; they are not derived from the word “effortless”. See OpenStax, Viscosity and Turbulence.

Direct connection to effort = force. Let an ideal piston with fixed area \(A_p\) deliver the same total throughput. The required pressure contribution to the driving force is \(F=A_p\Delta p\). With both tubes open, \(R_{\rm eq}=2R_A/3\). If we close only the longer tube B, then \(R_{\rm eq}=R_A\), so that in the new stationary state: \[ \boxed{\frac{F_{\rm B\ closed}}{F_{\rm both\ open}}=\frac32.} \] The same displaced quantity per unit time therefore requires 50% more piston force. Piston friction, connection losses and acceleration during switching are not included here. At fixed pressure instead of fixed throughput, the total flow would decrease; that is a different experiment.

Reading within EEML: both routes contribute to the same effect, with the largest contribution via the route with the lowest resistance. The longer route nevertheless lowers the required driving force of the whole. Under these conditions, “everything exclusively takes the easiest way” is therefore incorrect. At fixed total throughput, the quadratic loss \(P=R_AQ_A^2+R_BQ_B^2\) has its minimum precisely at this distribution, as the network identity above shows.

Evidential strength: an existing law is correctly reconstructed under its own hypotheses and read in EEML terms. The 2:1 ratio and the force factor 3/2 are testable, but they do not thereby yet support EEML independently of Poiseuille and flow conservation. This benchmark does not investigate spontaneous tube growth, memory or the process law of P1.

#Electrical benchmark: the same charge current with less heat generation

Known retrodiction, added on 28 September 2026. Take two parallel, ohmic resistors at constant temperature: \(R_A=10\,\Omega\) and \(R_B=20\,\Omega\), connected to a source that keeps the total current at 1 A. Connection losses and the switching transient are left outside this ideal stationary example. The effect is the same net charge transport per second, over the same measurement duration.

Ohm and current conservation predict: \[ I_A=\frac23\,{\rm A},\quad I_B=\frac13\,{\rm A},\quad U=\frac{20}{3}\,{\rm V},\quad R_{\rm eq}=\frac{20}{3}\,\Omega. \] The heat generation is: \[ P=R_AI_A^2+R_BI_B^2=\frac{20}{3}\,{\rm W} \simeq6.67\,{\rm W}. \] With only route A open, the same effect requires \(U=10\,{\rm V}\) and \(P=10\,{\rm W}\). The extra route therefore lowers the required voltage and heat generation by one third. Over 60 seconds, the conversion into heat is 400 J versus 600 J, with 60 C of net charge passed through in both cases. See OpenStax, Resistors in Series and Parallel.

The minimum follows exactly: for \(I_A=2/3\,{\rm A}+\delta I\) and \(I_B=1/3\,{\rm A}-\delta I\), we have \[ P-P_{\min}=30\,\Omega\,(\delta I)^2\ge0. \] This compares current distributions on the same network at the same total current. It does not prove that nature designed the network itself or spontaneously lays down a new wire.

Particles and force: in metals, electrons carry the current as a net drift. Their electric force is \(\mathbf F=-e\mathbf E\). Current, voltage and force are different quantities: voltage is energy per charge, while the local field determines the force. The energy flow is accounted for here as electrical power and converted into heat in the resistors; energy does not disappear. A current distribution is no proof that individual electrons plan ahead. See Feynman, Linear Systems and Review.

Condition that determines the outcome: at a fixed voltage of 10 V, adding B would instead raise the current from 1 to 1.5 A and the power from 10 to 15 W. Then the effect is not held equal. The energy saving of one third belongs explicitly to the comparison at fixed total current.

Status for EEML: an exactly known case of less electrical drive and less dissipation for the same effect. The local laws are input. This is a concrete benchmark for EEML and not an independent derivation of Ohm, electron physics or a universal minimum principle.

#Zero active resistance without empty structure

For a symmetric positive semidefinite damping matrix Γ:

Γ need not be zero. With P = γ(v₁−v₂)², the channel does not slow down a joint motion v₁ = v₂, but it does slow down a relative motion. This preserves a substantive part of the zero-point idea without presupposing an origin of the universe.

A cost shift S → S + c creates a zero on a score axis; it does not switch off any physical dissipation channel. Likewise, ε → 0 is a different operation. The operational zero point, the zero point of an accounting cost and the sharp-selection limit remain separate.

Under its conditions, the older thermal realization K1 connects motion and weight:

For constant J,D, suitable boundary conditions and finite Z, e^(−H/ε)/Z is invariant. H, mobilities and thermal context remain input; uniqueness and attraction require more. Conservative motion can continue without active dissipation, but the same class can also allow rest. See the related matrix framework of Ma, Chen and Fox.

#Habits and structural memory

EEML-T 0.3 uses route shares x, learned preferences h and the current target distribution p, each on the probability simplex. The parameters κ, η and ε are positive; Φ is a differentiable potential of x. The chosen adaptation law reads:

With \(D_{\rm KL}(x\|h)=\sum_i x_i\log(x_i/h_i)\), positive shares, fixed conditions and Φ independent of h and time, Ψ = Φ + εD_KL(x‖h) gives:

Under an imposed blockage, the conditions change; Ψ may then jump between phases. The decrease holds within each phase that respects the hypotheses. A concrete restriction to open routes, with memory retained over all routes, is worked out in P1; an arbitrary renormalization or change of costs does not automatically fall under the proof. This is a precise example of “use changes what is effortless later”. Changing preference is not the same as changing physical resistance. Fixed preference, learning preference and changing conductance must be compared separately. The formula is model input with a derived consequence, not yet a validated universal learning law.

Choosing the effect in advance and comparing only successful outcomes is useful for design. But it does not predict how often success occurs. For that, physically possible failures must remain in the probability space.

For trajectory space Γ_C and success event E:

When a justified positive path weighting is used, this can be written as Z_E,C/Z_C. The path measure and any path cost must then have been specified independently. Normalizing only over Γ_E,C gives P(E|E,C) = 1, not the unknown success probability. With four equally weighted paths of which one succeeds, that probability is 1/4.

The following control example makes the difference concrete. Take A ↔ B with rate k in both directions. All k > 0 have the same stationary occupation (1/2,1/2). Start in A. Success means: the first arrival in B no later than T, then remaining there without interruption for Δ. A failed first arrival is not repaired by later attempts. Then:

For T = 1 and Δ = 0.5:

Moving fast helps with arrival, but here it harms the subsequent stay. The maximum within this family lies at k = ln((T+Δ)/Δ)/T. That is a mathematical optimum for the chosen criterion; no mechanism has been added that makes k move to k by itself.

Why this must return: equal weights and even the same possible transitions do not yet give an equal probability of a usable effect that is sustained for long enough. The latest statistical layer alone would miss this difference.

The following four corrections delimit the statistical formulation. MPP stands for the Multiple Point Principle, discussed in Part IV.

#The convexity theorem holds under fixed conditions

Take a fixed space, a λ-independent measure μ, fixed ε > 0 and Sλ = S₀ + λI. If Z is finite and may be differentiated twice under the integral:

With a flat macro-measure on a closed bounded interval, a maximum of Z is attained at an endpoint; an interior stationary point is a minimum, possibly not strict. The crossing example in Part IV is a concrete illustration of this.

The conclusion “MPP can come only from ν in this formalism” is, however, too general. For nonlinear parameter dependence, with otherwise the same conditions:

The first term can be positive. A simple counterexample is S₁ = λ², S₂ = 2λ² on [−1,1]. At λ = 0 the minima are equal and Z = e^(−λ²/ε) + e^(−2λ²/ε) is maximal, with flat ν. F''(0) = 3. This does not prove a physical MPP, but it refutes the universal exclusion of a preference from Z. A changing measure or support can also invalidate the affine theorem.

The recorded position remains: MPP does not follow from the current core; a separately justified parameter selection mechanism is needed. The affine canonical route with fixed measure has the above limitation.

#Refinement: genuine coefficients, parameter maps and a negative control

The additional review rightly notes that \(S_1=\lambda^2,\ S_2=2\lambda^2\) becomes affine in the coefficient \(g=\lambda^2\). On \(g\in[0,1]\), the preferred value \(g=0\) is a boundary point. This example therefore does not refute the affine coefficient theorem. It does, however, limit statements about arbitrary parameter maps. That a mathematical reparametrization is possible does not by itself prove which parameter or measure is physically fundamental.

For fixed space and measure, finite normalization and \(S_g=S_0+\sum_a g_a I_a\), the following holds:

The Hessian is positive semidefinite. On a compact convex coefficient range, a maximum can also be found at an extreme point; in flat cases, interior points can also be maximizers. This is not a statement about all coupled, nonconvex parameter sets, parameter-dependent measures or infinite partition functions.

Negative control: one positive cost contribution. If \(S_g=S_0+gI\), \(I\ge0\), \(g\ge0\), and the fixed measure and differentiation conditions hold, then:

With finite Z at g=0 and a flat macro-measure in g, the density is maximized there. The finite example \(S=(0,g)\), ε=1, gives \(Z(g)=1+e^{-g}\). If an application independently needs a nonzero g, this bare rule therefore does not select it. This is a negative model control, not a new measurement.

For \(I\le0\), the same formula instead gives a nondecreasing Z. ‘The maximum lies on a boundary’ requires that a suitable maximum exists: for example on a closed bounded interval with finite continuous Z. For an operator without fixed sign, \(\langle I\rangle\) can change sign; with fixed measure and affine costs, log Z remains convex, so that an interior stationary point is not a strict maximum.

On an unbounded or open range, there may be only a supremum at a missing boundary, or the normalization over couplings may diverge. For example, \(Z(g)=1+e^{-g}\) is finite for every \(g\ge0\), but \(\int_0^\infty Z(g)\,dg=\infty\). A flat macro-measure therefore yields no normalized probability distribution at all there. If the effect criterion or the allowed set itself depends on g, the derivative at fixed space does not automatically hold either. A non-flat ν can change the probability density without changing Z.

It does not follow from this that all canonically weighted theories are free and symmetric. Other interactions can remain in \(S_0\); a zero coefficient can give a divergent Z; and zero in an action coefficient is not automatically zero in every customary physical coupling convention. The regularized positive \(\phi^4\) contribution is a suitable example only if the other assumptions, including normalization at the boundary, also hold.

The macro-measure remains essential. A flat measure on \(\lambda\in[-1,1]\) becomes, under \(g=\lambda^2\) with both branches included, proportional to \(g^{-1/2}dg\), not to dg. Silently taking a flat measure in both coordinates changes the model. Moreover, for a continuous distribution, a highest density is not a probability of one at that point.

Decision for the core: \(P(dg)\propto\nu(dg)Z_g\) remains a possible, conditional macro-ensemble description. It does not become a general law of nature that chooses couplings. An effect requirement can restrict the allowed range, but does not by itself determine a single nonzero coupling. The measure, the effect and the selection process must be justified independently.

#Width is not the same as instability

A positive quadratic minimum S = κx²/2 is stable under the chosen dynamics ẋ = −κx for every κ > 0. Small κ gives a wide minimum and slow return, not negative stability. Escape under noise also requires barriers, mobility and an observation time.

There can be a trade-off between statistical weight, recovery rate and reliability. There is no necessary conflict between wide minima and stability. Therefore, use separate quantities for curvature, return time, escape probability and measure volume. The control example confirms this for κ = 0.01 and κ = 100.

#Noncompact moduli do not automatically mean infinite measure

A noncompact space can have finite or infinite measure. Even with S = 0 on the whole real line, ∫dx diverges, while ∫dx/[π(1+x²)] = 1. The cost and the space are the same; the measures differ.

Therefore, “infinite basin measure” does not follow merely from the number of K3 moduli or from noncompactness. For that, the precise space, quotients, volume treatment, measure and any potential must be specified. The Cauchy measure from the example is not proposed as a physical K3 measure. Normalizability remains a condition to be checked, not an impossibility proven in advance.

#Measure volume, reachability and efficiency remain separate

“Largest measure-weighted accessible volume” can describe statistical weight. It is not automatically maximal efficiency, and “accessible” does not follow from an equilibrium integral. The actual initial preparation and transition law determine whether that volume is reached within the relevant time. An ordinary generator K or L, too, contains memory only if the state has been extended sufficiently for it.

The original ideas remain visible: nothing does not exist; effortless effect as the underlying idea; motivations, habits and goals as context; self-application; and the convention that “100% effortless” means zero active opposition with actual effect. Continuing operation remains a proposed universal property. These statements are not a consequence of the partition function and need not silently disappear in order to use the concrete models.

Not returning as a proven core:

• “Zero resistance necessarily causes motion.” A drive or other dynamical structure remains necessary; zero active dissipation also allows rest.
• “Everything always reaches the global minimum” or “every standstill is unstable”. Existing model counterexamples still stand.
• “Existence is positive resistance” as a necessary and sufficient condition; this has not been shown and does not fit the retained zero-active-resistance states.
• A universal conservation of resistance, a unique resistance density per domain or a self-proven unique meta-law.
• Topological stationarity as moduli stabilization: a topological invariant is constant within a fixed sector and does not choose a volume or coupling there.
• Multiple minima as a derivation of superposition, or a “zero-effort limit” as a derivation of the zero point.
• The old 137, 18, Higgs, dark-sector and LISA numbers without the missing valid calculations.

K3 and the particle benchmarks remain instructive case studies with their checked conditions. They do not determine the general architecture in advance.

The most concrete connection is EEML-W for executability and support + a tested process law + the gluing rule for the same observations at multiple scales. EEML-T 0.3 and K1 are different available model realizations; they must not be merged into one universal equation without derivation.

The appropriate research direction is: reconstruct transition counts and residence times, and then predict a new blockage. Besides occupation, the next physical test must also measure events, waiting times and support. For a new EEML claim, one then needs to know which extra relation EEML predicts on top of the reference model.

Retained core: carrying effect, continuing operation, comparing required input and support, safeguarding coherence and predicting a new consequence. Open heart: which independently motivated relation between these components nature actually follows and existing separate domain models do not yet impose.