# Observers, Blankets, and Screens: Executable Correspondences, Obstructions, and Complementarities across Observer Patch Holography, the Free Energy Principle, and Fuller’s Synergetics

> Literal text extraction from `OPH-FEP-Synergetics_combined.pdf`; SHA-256 `7409d8d1a9cf3073da0712781cde0e209a139bee5c5dccde85af3795540592a3`. PyMuPDF 1.28.2; no OCR or scientific validation.


## Page 1

Observers, Blankets, and Screens
Executable Correspondences, Obstructions, and Complementarities across Observer Patch
Holography, the Free Energy Principle, and Fuller’s Synergetics
Daniel Ari Friedman
Active Inference Institute
daniel@activeinference.institute
ORCID: 0000-0001-6232-9096
November 15, 2026


## Page 2

1
Graphical abstract: three pairwise ledgers in four panels
Every claim in this paper is one correspondence: a statement about a single object, put in two or three vocabularies at
once, executed by a probe that returns values rather than a verdict. The registry holds 44 such claims: 19 identifications, 15
obstructions and 10 complementarities, scoped to the pairs oph-fep (26), synergetics-oph (14), fep-synergetics (2) and
three-way (2).
The four panels are the paper in pictures, one page each: the ledger of the three pairwise scopes; the carrier every scope
is about; the dynamics that separates the readings; and the accounting, where the programmes disagree about currencies
rather than facts.
1.1
Panel 1 — the pairwise ledger: what each pair of programmes agrees on, obstructs, and
supplies
Each edge carries one pair’s ledger in the form the command line executes it: oph-fep-synergetics verify --pair
oph-fep, --pair synergetics-oph, --pair fep-synergetics. The centre carries the one claim that needs all three vocab-
ularies at once. Most OPH–FEP work is obstruction: 26 claims, most of them failures of an identification proposed in good
faith. Most Synergetics–OPH work is identification: 14 claims about symmetry, volume and counting. The FEP–Synergetics
edge is short by 2 claims and holds two exact results about information, in one of which the programmes disagree about the
meaning of a number.
Figure 1: The pairwise ledger triangle. Vertices are the three programmes, Observer Patch Holography, the Free Energy
Principle and Fuller’s Synergetics. Each edge is annotated with the counts of identifications, obstructions and complemen-
tarities for that pair, and the three load-bearing identifications are named on the edge where they sit. The centre carries the
three-way ledger: the two claims that need all three vocabularies at once, tetravolume-arithmetic – three exact quantities
in three number systems – and conjugation-norm. Every count is read from the executed registry at draw time rather than
transcribed.
The scope of a claim is part of the claim: carrier-port-chart is a statement about OPH’s chart and the icosahedron’s
vertex set, spectral-gap is a statement about one matrix read two ways — the generator of OPH’s mean repair operator
and the FEP’s precision — with the gap inherited from the shared object, and kl-lattice-metric is a statement about the
FEP’s divergence and Synergetics’ lattice metric. sec. 6 reports the three ledgers separately, and tbl. 4 carries the scope of
every row.


## Page 3

1.2
Panel 2 — the carrier: one graph, its symmetry, its counting identities, and its Gaussian
reading
Every scope is read on the same object, so the paper can compare vocabularies without first having to reconcile subjects:
twelve ports, thirty seams, twenty triangular faces, every port of degree 5, spectral gap 2.763932. OPH supplies the repair
law on the seams of that graph. The FEP reads a diagonal shift of the Laplacian as a proper Gaussian precision. Two ports
in this model are dependent given the rest exactly when they share a seam. The panel also carries two counting identities
that neither programme states but both inherit from the graph: the number of spanning trees, and the dimensions of the
agreement sheaf (Proposition 7; Proposition 9).
Figure 2: The carrier and its exact invariants. Left: the ports and seams drawn from their derived adjacency. Right: the
spanning-tree count, sheaf dimensions, and symmetry counts, followed by centered intrinsic port variance and the proper
Gaussian’s blanket size and pairwise precision.
The reduced Laplacian of the seam graph has determinant 5184000, so the carrier has exactly that many spanning trees,
and the same number is recovered from the spectrum as a product of eigenvalues. The scalar sheaf on the seam graph has a
1-dimensional space of global sections and 19 independent obstructions, the coboundary has rank 11, and the graph’s Euler
characteristic is -18, the twelve nodes minus the thirty edges; with the twenty faces added their boundaries span the cycle
space, and the surface’s Euler characteristic is 2. The carrier’s automorphism group has order 120, splits into 60 rotations
and 60 reflections when the orientation of OPH’s own faces is used as the test, and acts on the 78 total-two configurations in
4 orbits with 3 distinct descent-outcome multisets. The proper Gaussian has seam precision -1 and a blanket of 5 neighbours
per port. The centred intrinsic field gives port variance 0.1944444, exactly 7/36.
3


## Page 4

1.3
Panel 3 — native repair and Gaussian costs on lattice shells
Native repair takes a configuration, a seam and a coin. Of the 78 states in its census, 66 are binary. Over 4680 attempts,
120 lower the load energy, 600 move with the energy unchanged, and 3960 leave the state unchanged.
The preceding
panel’s equivariance explains why this census is invariant under carrier symmetry. The separate lattice example compares
equal-covariance Gaussian divergence with the IVM shell index.
Figure 3: The descent census and the Gaussian divergence levels. Left: strict descents, zero descents and no-ops on the
native load shell, with the path multiplicities. Right: half the squared norm over the first five IVM coordination shells, one
column per shell, with shared levels marked.
One lattice step costs 4 nats at unit covariance, because the Gaussian divergence between equal-covariance neighbours is half
the squared norm and the first shell sits at squared distance 8, so the sites at squared norm 8𝑘2 cost 4𝑘2. The first five shells
carry (1, 3, 4, 7, 7) distinct free-energy levels, and the levels (36, 52) are each carried by two shells: the shell index does not
fix the free energy, and the free energy does not fix the shell. The shell is a level set of a polyhedral norm and the free energy
of a Euclidean one, and the obstruction is that neither refines the other.
4


## Page 5

1.4
Panel 4 — the accounting: tetravolumes, one packing, and three number systems
Synergetics counts volume in tetrahedra of the prime vector and asks which solids come out integral. The IVM’s own solids
do: (1, 3, 4, 6, 20, 96). The carrier’s own solid does not: it enters at 18.5123 tetravolumes, which is (5/2)(3
√
2 +
√
10)
exactly. The same accounting settles the twelve-around-one question completely, because the cell of the twelve neighbours is
a rhombic dodecahedron whose volume can be derived rather than looked up, and the ball inscribed in it gives the densest
packing density in three dimensions [Hales, 2005].
Figure 4: Tetravolume accounting and the twelve-around-one cell. Left: the tetravolumes of the IVM’s integral solids beside
the regular icosahedron’s measured tetravolume on a logarithmic axis, showing that the carrier’s own solid is the one entry
an integer table cannot express. Right: the Voronoi cell of the twelve neighbours with its vertex, face and edge counts, its
derived tetravolume from half the squared norm at each of its vertices, and the packing density of the inscribed ball against
the closed form.
The cell of the twelve neighbours has 14 vertices, 12 faces and 24 edges, tetravolume 6 derived from the halved squared norm
at each of its vertices, and an inscribed ball of 4.442883 tetravolumes which fills density 0.7404805, exactly 𝜋
√
2/6 = 𝜋/(3
√
2)
because the cell’s volume is rational and the ball’s tetravolume is 𝜋
√
2. The carrier’s icosahedron enters at 18.5123, computed
exactly as
√
2 times an element of ℚ(
√
5), with squared value 342.7051, exactly 175+75
√
5, and is not integral: the tetravolume
table and the carrier’s own solid live in different arithmetics even though they are the same geometry. The three programmes’
exact quantities live in three number systems: an integer 4 nats for the FEP’s step, the quadratic surd 0.04606553 for OPH’s
relaxation rate, and integral tetravolumes beside one biquadratic one in Synergetics’.
5


## Page 6

Contents
1
Graphical abstract: three pairwise ledgers in four panels
2
1.1
Panel 1 — the pairwise ledger: what each pair of programmes agrees on, obstructs, and supplies
. . . . . . .
2
1.2
Panel 2 — the carrier: one graph, its symmetry, its counting identities, and its Gaussian reading . . . . . . .
3
1.3
Panel 3 — native repair and Gaussian costs on lattice shells . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4
1.4
Panel 4 — the accounting: tetravolumes, one packing, and three number systems . . . . . . . . . . . . . . . .
5
2
Abstract: three programmes, one carrier, and a registry that executes every claim
8
3
Introduction: three programmes of observer, one finite carrier, and a decision procedure for identifi-
cation
9
3.1
Three programs, one question . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
9
3.2
What a correspondence is . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
9
3.3
Contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
9
3.4
What this paper does not do
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
10
3.5
Related work and positioning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
10
4
The three programs and the one object they share, stated in their own vocabularies
11
4.1
The carrier
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
11
4.2
OPH: a repair law on the seams . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
11
4.3
FEP: a bound on surprisal . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
12
4.4
Synergetics: tetravolume accounting
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
13
4.5
One object, three vocabularies, and a dictionary this paper tests
. . . . . . . . . . . . . . . . . . . . . . . . .
14
5
Method: a registry whose every claim is executed, scoped and re-runnable
15
5.1
The unit of work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
15
5.2
The registry, grouped by pairwise scope
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
15
5.3
Reading a source without copying it
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
18
5.4
Why exactness, not agreement
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
19
6
Agreement: the identifications each pairwise scope survives
20
6.1
The port chart is a consequence, not a transcription
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
20
6.2
The spectrum is closed-form . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
20
6.3
The shell is counted exactly . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
21
6.4
The native repair law is reproduced, not approximated . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
21
6.5
Each native repair is exact conditioning
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
22
6.6
The free-energy dictionary has a precise scope . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
23
6.7
A sheaf computes the same Laplacian
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
23
6.8
The carrier counts its own spanning trees
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
24
6.9
The Gaussian reading makes the seam graph the independence graph . . . . . . . . . . . . . . . . . . . . . . .
24
6.10 Two cohomology dimensions, one global section . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
24
6.11 The symmetry group closes the census . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
25
6.12 Information cost is the synergetic metric over two . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
25
6.13 The carrier’s subgraphs are the solid’s own counts
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
26
6.14 Curvature is exact in units of pi . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
26
6.15 What the agreements do not say
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
26
7
Disagreement: the obstructions, each stated exactly
28
7.1
The native law does not terminate
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
28
7.2
Three distinct values at fixed records . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
29
7.3
A stationary blanket does not bound the dynamics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
29
7.4
Stationary blankets leave entropy production free . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
29
7.5
Stability is not generic in the coupling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
30
7.6
Reduction needs a threshold . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
31
7.7
Existence does not imply reachability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
32
7.8
Direct and adjoint gradients differ
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
32
7.9
The four-cycle has an incompatible local minimum . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
33
7.10 Twelve vertices is not one solid
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
33
6


## Page 7

7.11 The shell index and the free energy determine neither each other . . . . . . . . . . . . . . . . . . . . . . . . .
34
7.12 The shells around a port are not the lattice’s shells . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
34
7.13 Twelve around one names a family, not an object . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
34
7.14 What the obstructions have in common
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
36
8
Complementarity: structures one programme supplies and another needs
37
8.1
Synergetics supplies the accounting in which the carrier’s geometry is exact . . . . . . . . . . . . . . . . . . .
37
8.2
The same quadratic sequence appears three times . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
38
8.3
Energy-based learning supplies the gradient the repair law lacks . . . . . . . . . . . . . . . . . . . . . . . . . .
38
8.4
Selection supplies a potential flow, for symmetric payoffs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
39
8.5
The carrier’s own solid has an exact algebraic volume
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
39
8.6
The densest packing in three dimensions has the twelve-around-one cell
. . . . . . . . . . . . . . . . . . . . .
40
8.7
Three number systems and no common unit . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
40
8.8
The precision also carries a resistance metric
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
41
8.9
The twelve-around-one cell and the cuboctahedron are duals . . . . . . . . . . . . . . . . . . . . . . . . . . . .
41
8.10 Conjugation fixes the counts and moves the volumes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
41
8.11 What the complements add up to . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
41
9
Scope, limits and falsifiers of a correspondence battery
42
9.1
What is verified . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
42
9.2
Three source quantities and their protocols
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
43
9.3
What is not claimed
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
43
9.4
Four cautions for anyone continuing this work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
43
9.5
What would change the conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
43
10 Reproducibility: pins, commands, and the binding of every number to a function
45
10.1 What this manuscript is made of . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
45
10.2 The three source pins . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
45
10.3 The commands . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
45
10.4 Environment
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
46
10.5 What a reader should check first
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
46
11 References
47
7


## Page 8

2
Abstract: three programmes, one carrier, and a registry that executes
every claim
Three research programs describe observers in incompatible vocabularies. Observer Patch Holography builds spacetime from
a finite complex of patches joined by seams, with a local repair law that moves load across those seams [FloatingPragma, 2026,
Mueller et al., 2026]. The Free Energy Principle derives the behaviour of a self-maintaining system from a variational bound on
surprisal [Friston, 2010, Parr et al., 2022]. Fuller’s Synergetics axiomatises a closest-packed, tetravolume-counted geometry in
which the twelve-around-one configuration is the unit of spatial accounting [Fuller and Applewhite, 1975, Edmondson, 1987].
This paper reports what happens when the three are made executable against one another. A registry of 44 correspondences
was built, each one a claim scoped to a pair of programmes and carrying a probe that either reproduces an exact value or
records the failure precisely: 19 isomorphisms, 15 disagreements and 10 complementarities. By scope, 26 claims relate OPH
to the FEP, 14 relate Synergetics to OPH, 2 relate the FEP to Synergetics, and 2 entries need all three vocabularies at
once. The load-bearing agreements are structural. OPH’s carrier is a 12-port, 30-seam, 20-face complex whose adjacency is
recovered from its own oriented faces, and its Laplacian spectrum is closed-form, with spectral gap 2.763932. The native
repair law conserves load and lowers a quadratic energy by a closed-form amount, and over the 78-configuration load shell
it produces 120 strict descents, 600 zero descents and 3960 no-op moves out of 4680 transitions. On the binary shell each
native repair is exact Bayesian conditioning on the datum it leaves visible, and its law minimises a variational free energy
whose minimum is the surprisal of that datum [Mueller, 2026, Mueller et al., 2026].
Counting the same object turns out to add identifications that no dynamical reading supplies. The seam graph has 5184000
spanning trees, an integer that survives the surds in its own spectrum; the scalar sheaf on the seam graph has a 1-dimensional
space of global sections (the constants) and 19 obstructions, the cycle rank. The face boundaries span the cycle space, giving
surface Betti numbers (1, 0, 1).
The automorphism group has order 120, splitting into 60 rotations and 60 reflections,
transitive on ports, seams and faces, with the descent census invariant across all 78 configurations and 4 orbits under it. Its
subgraphs are counted the same way: 20 triangles, exactly the faces the port chart was derived from, 12 five-cycles for the dual
solid’s vertices, 125 perfect matchings, and 1280 Hamiltonian cycles, with a curvature of 4 pi that is Gauss-Bonnet exactly
and a resistance ladder whose exact fractions sum to a Kirchhoff index of 28. A proper Gaussian with diagonally shifted
Laplacian precision has the seam graph as its conditional-dependence graph, with a Markov blanket of 5 neighbours per
port. Separately, equal-covariance Gaussian divergence makes one lattice step cost 4 nats at unit covariance, half its squared
distance in Synergetics’ metric. Synergetics’ accounting, every entry of which is derived here exactly from an explicit vertex
set, then extends to the carrier’s solid, which measures 18.5123 tetravolumes, an exact element of
√
2 ℚ(
√
5), against the
integral table (1, 3, 4, 6, 20, 96), and to the twelve-around-one cell, whose 14 vertices, 12 faces and 24 edges give tetravolume
6 and the packing density 0.7404805, exactly 𝜋/(3
√
2).
The load-bearing disagreements are equally structural. On the tested shell the native law is confluent but does not terminate;
the broken-seam gate terminates without giving unique normal forms. The binary-shell paths preserve load energy, while the
ensemble can lose relative entropy. A pairwise reduction of a boundary exists only above a mediator threshold 0.3333333; a
blanket in the stationary density does not constrain the dynamics that preserve it; the solenoidal family leaves a stationary
blanket’s entropy production free [Mueller, 2026]; and four of the obstructions are failures of counting rather than of dynamics:
twelve vertices do not determine a solid, a coordination shell and a free-energy level are not functions of one another, a
neighbourhood is not a lattice shell, and the phrase “twelve around one” does not determine which solid is meant. Synergetics
contributes the one geometry in which the other two can be drawn, not a derivation of either.
Three quantities the source literature states for the native law are reproduced exactly from OPH’s public model, and the
scope limits are stated rather than trimmed. Every number in this manuscript is bound to the function that produces it,
every claim is executed by oph-fep-synergetics verify (or narrowed with --pair and --kind), and the graphical abstract
on the preceding pages states the same results as four drawn panels.
8


## Page 9

3
Introduction: three programmes of observer, one finite carrier, and a de-
cision procedure for identification
3.1
Three programs, one question
Observer Patch Holography (OPH) is a discrete reconstruction programme: a finite complex of patches carries integer load,
seams between patches carry a repair law, and geometry is what the repair statistics look like from inside [FloatingPragma,
2026, Mueller et al., 2026, Friedman, 2026a]. The Free Energy Principle (FEP) is a variational programme: a system that
persists must keep its sensory surprisal bounded, and the bound is a functional of a recognition density [Friston, 2010, 2013,
Parr et al., 2022]. Synergetics is a geometric programme: spatial accounting is done in tetravolumes on a closest-packed
lattice generated by twelve equal spheres around a central one, and the lattice’s own sites and solids are counted with integers
[Fuller and Applewhite, 1975, Friedman, 2025].
Each program answers a question the others only gesture at. OPH supplies a finite, exactly specified substrate and an explicit
local update. FEP supplies the functional that a persisting system must be minimising, together with a statistical notion of
a boundary. Synergetics supplies a coordinate system in which the lattice’s finite configurations are counted without square
roots, and this paper finds the one place where that fails: the carrier’s own solid.
The question this paper asks is narrow enough to be answerable: where the three programs make claims about the same finite
object, do those claims agree? Agreement here means more than analogy. Two claims agree when they can be stated as
functions of the same object and the values coincide.
3.2
What a correspondence is
A correspondence is a record with four parts Definition 4: a claim, a pairwise scope, a probe, and a status. The claim names
what OPH, FEP and Synergetics each say. The scope names which two or three of them the claim relates, because a claim
that two vocabularies are about one object is a claim about those two. The probe is executable code that computes the values
the claim rests on. The status is one of four terms, used in a fixed sense throughout sec. 5.
exact means the probe reproduces the claimed value exactly, in integers or rationals, or to floating-point equality with the
closed form. exact-on-domain means the identification holds on a stated domain, typically the linear-Gaussian corner, and
the probe asserts both the identity on that domain and its failure off it. obstructed means the identification fails, and the
probe exhibits the failure with the same precision as a success. not-reproduced means the source asserts a quantity that its
text does not determine, and the probe records the gap rather than inventing a value.
This vocabulary is the paper’s main methodological commitment. An obstructed correspondence is a result, not a defect. It
says where the three programmes cannot be welded together, and therefore where a synthesis has to choose.
3.3
Contributions
The contributions are the following.
1. A derived carrier. OPH’s twelve-port adjacency is recovered from its twenty oriented faces rather than transcribed from
a table, and the derivation is checked against an independent OPH consequence sec. 4.
2. A closed-form spectrum and gap for that carrier, the gap a quadratic surd and the multiplicities forced by the icosahedral
group Proposition 2.
3. A complete census of the native repair law over the 78 configurations of total load two: 120 strict descents, 600 zero
descents, 3960 no-op moves Proposition 14. The reachability probe separates the ungated law’s nontermination from
the gated law’s lack of confluence Proposition 15.
4. An executable reproduction of Mueller’s linear-Gaussian equivalence [Mueller, 2026, theorem 7, items 2–3]. The package
checks the harmonic extension and the limit of predictive-coding perception against a separate KKT solve on the carrier
Proposition 5.
5. Executable reproductions of Mueller’s mediator, solenoidal and stability controls [Mueller, 2026, propositions 8–10 and
15]. The package derives the general Routh column and onset 3.162278 Proposition 18, verifies that the decomposition
at the mediator threshold 0.3333333 uses rank-one blocks Proposition 19, and checks the solenoidal entropy production
Proposition 17. Exact enumeration of the clipping regimes and an independent orbit measurement extend the two-node
control Proposition 20.
6. An exact correspondence with Synergetics’ IVM: twelve neighbour moves at squared distance 8, coordination shells of
10𝑘2 + 2 sites identified in closed form as the lattice points on the frequency-𝑘vector equilibrium, and the tetravolume
table derived exactly from explicit vertex sets Proposition 4.
9


## Page 10

7. Two counting identities for the same carrier that no dynamical reading supplies: 5184000 spanning trees, and a sheaf
whose 19 graph obstructions are exactly the face boundaries, so that the carrier’s surface has Betti numbers (1, 0, 1)
Proposition 8, Proposition 9.
8. The carrier’s symmetry, which is what makes the census an invariant rather than a coincidence: order 120 split into
60 rotations and 60 reflections, irreducible eigenspaces, and a repair law equivariant transition by transition over 78
configurations in 4 orbits Proposition 10.
9. One place where the FEP’s currency and Synergetics’ metric are the same function: an equal-covariance divergence of
4 nats per lattice step at unit covariance, so a site at squared norm 8𝑘2 costs 4𝑘2 Proposition 11.
10. An exact volume for the carrier’s own solid and for the twelve-around-one cell, both computed without floating point:
18.5123 tetravolumes,
√
2 times an element of ℚ(
√
5) whose square is 175 + 75
√
5, against an integral table, and a
packing density of 0.7404805, exactly 𝜋/(3
√
2), derived from the cell rather than quoted Proposition 29, Proposition
30.
11. Five readings of the carrier as a network, none of which either programme supplied: exact effective resistances summing
to a Kirchhoff index of 28, a subgraph census of 20 triangles, 12 five-cycles, 125 matchings and 1280 Hamiltonian cycles,
and a curvature of 4 pi that is Gauss-Bonnet exactly Proposition 32, Proposition 12, Proposition 13.
12. A provenance discipline in which every source claim is pinned to a commit and every reimplementation cites the identifier
it reimplements sec. 10.
Items 1 to 11 are executable: each is a function, and each section below reports the values that function returns.
3.4
What this paper does not do
It does not claim that OPH is a free energy principle, or that FEP is a holographic theory, or that Synergetics predicts either.
It tests identifications from the integration proposals and runs the OPH–FEP controls developed in Agreement and Surprise
[Friedman, 2026a,c, Mueller, 2026]. It reproduces the three quantities the source literature states for the native law from the
protocols in OPH’s public model sec. 9. It does not compare any of the three against laboratory data; OPH’s own empirical
status document is explicit that no retained comparison selects its record law, and this paper inherits that situation rather
than resolving it [FloatingPragma, 2026].
3.5
Related work and positioning
The FEP literature contains its own internal disagreements, and this paper uses them as evidence rather than as noise.
The Markov blanket is a statistical conditional-independence structure [Pearl, 1988], not automatically a dynamical object
[Bruineberg et al., 2022], and technical objections to parts of the free energy principle have been stated precisely [Biehl et al.,
2021]. The variational free energy is a bound, and its exactness conditions matter [Friston and Stephan, 2007, Sajid et al.,
2021]. Bayesian mechanics gives the stationary-process reading of the same mathematics [Da Costa et al., 2021, Parr et al.,
2020], and the discrete-state synthesis supplies the notation for the active-inference corner [Da Costa et al., 2020].
The title’s third noun has a precedent on the FEP side: a Markov blanket has been read as a holographic screen, the
boundary through which a system and its environment exchange information [Fields et al., 2022]. OPH’s patches are screens
in a different, discrete sense, and this paper does not identify the two; it records where their finite shadows agree. Bayesian
mechanics has also been criticised from the physics side for how much of a system’s dynamics a stationary blanket actually
fixes [Aguilera et al., 2022], which is the question sec. 7.3 makes exact.
Synergetics has a small technical literature that states its conventions carefully [Edmondson, 1987], and the geodesic sub-
division it motivates is the Goldberg construction [Goldberg, 1937] on the regular icosahedron [Coxeter, 1973], the same
construction that classifies icosahedral virus shells [Caspar and Klug, 1962]. The lattice that results is the root lattice A3 in
the closest-packing literature [Conway and Sloane, 1999], whose kissing number twelve was established by Schütte and van
der Waerden [Schütte and van der Waerden, 1953] and whose coordination sequence is 10𝑘2 + 2 [Conway and Sloane, 1997].
The tools this paper uses have long pedigrees, and it claims none of them: the algebraic connectivity of a graph [Fiedler, 1973],
the Routh–Hurwitz criterion for a cubic [Routh, 1877, Hurwitz, 1895], Lyapunov stability [Khalil, 2002], contraction analysis
[Lohmiller and Slotine, 1998], energy landscapes in neural networks [Hopfield, 1982], the information measures the surprises
are built from [Shannon, 1948, Kullback and Leibler, 1951, Amari, 2016], topological signal processing [Ghrist, 2014], the
Markov-blanket literature [Kirchhoff et al., 2018], and the predictive-processing reading of the free energy principle [Clark,
2013, Friston et al., 2017, Friston, 2019].
The contribution here is not a new theory but a decision procedure: a correspondence either executes or it does not, and the
paper reports which.
10


## Page 11

4
The three programs and the one object they share, stated in their own
vocabularies
4.1
The carrier
All three programs, read charitably, make claims about one finite object: a twelve-vertex, thirty-edge, twenty-face complex.
This section fixes that object and then states each program’s mathematics on it.
Definition 1. A patch complex is a finite simplicial complex 𝐾= (𝑉, 𝐸, 𝐹) with twelve vertices (ports), thirty edges
(seams) and twenty triangular faces. A load configuration is a map 𝑥∶𝑉→ℤ; the total load is ∑𝑣𝑥𝑣. A load configuration
is binary when 𝑥𝑣∈{0, 1} for every port. The carrier is the specific complex whose twenty oriented faces are those of the
regular icosahedron under the port chart of [FloatingPragma, 2026].
The port chart is not decoration. OPH ships the twenty oriented faces, and the adjacency is a consequence of them: two
ports are joined exactly when some face contains both. This package derives the adjacency that way and then asserts that the
twenty triangles of the derived graph are exactly the twenty faces it started from, so a transcription error in either direction
would be caught.
Proposition 1. The graph obtained from OPH’s twenty oriented faces is 5-regular on 12 ports with 30 seams, and its 20
triangles are exactly the twenty generating faces.
The adjacency that results is listed in tbl. 1; it is reproduced here because the labelling differs from the common icosahedral
labelling in which port 0 meets ports 1, 2, 3, 4, 6, and a reader who assumes the common labelling will not reproduce any
downstream number.
Table 1: The derived carrier adjacency: every port meets 5 others, and the triangles of this graph are exactly OPH’s twenty
oriented faces.
Port
Adjacent ports
Port
Adjacent ports
0
1, 5, 7, 10, 11
6
2, 3, 7, 8, 10
1
0, 5, 7, 8, 9
7
0, 1, 6, 8, 10
2
3, 4, 6, 10, 11
8
1, 3, 6, 7, 9
3
2, 4, 6, 8, 9
9
1, 3, 4, 5, 8
4
2, 3, 5, 9, 11
10
0, 2, 6, 7, 11
5
0, 1, 4, 9, 11
11
0, 2, 4, 5, 10
The carrier’s Laplacian is the standard graph Laplacian 𝐿= 5𝐼−𝐴, which is how OPH writes it, and its spectrum is closed-
form. The algebraic connectivity is 5−
√
5 [Fiedler, 1973], and the gap normalised by the total degree 2|𝐸| is (5−
√
5)/60. In
OPH’s reading 𝐿is not the energy the repair law lowers (that is eq. 2 below); it is the generator of the mean repair operator
𝐼−𝐿/(2|𝐸|) and the quadratic form of the seam-disagreement energy 1
2 ∑{𝑎,𝑏}∈𝐸(𝑥𝑎−𝑥𝑏)2 = 1
2𝑥⊤𝐿𝑥.
Proposition 2. The carrier Laplacian has spectrum {0, (5 −
√
5)×3, 6×5, (5 +
√
5)×3}, so 𝜆1 = 5 −
√
5 ≈2.763932 and
𝜆1/(2|𝐸|) = (5 −
√
5)/60 = 0.04606553.
The multiplicity pattern is the one forced by the icosahedral symmetry. The permutation representation of the full icosahedral
group on the twelve ports splits into four pairwise inequivalent irreducibles, of dimensions one, three, five and three, so each
eigenvalue’s multiplicity is the dimension of one irreducible and no eigenspace contains two. The group fixes the eigenspaces
and their dimensions; the four eigenvalues themselves are particular to 5𝐼−𝐴Proposition 10. fig. 5 draws the carrier and
its spectrum.
4.2
OPH: a repair law on the seams
OPH’s native dynamics is a local rule that moves integer load across a seam. On a seam joining ports 𝑎and 𝑏, with 𝑑= 𝑥𝑎−𝑥𝑏
and a coin 𝑐∈{0, 1}, the amount moved is
𝛿= 𝑑−(2𝑐−1) (|𝑑| mod 2)
2
,
𝑥𝑎←𝑥𝑎−𝛿,
𝑥𝑏←𝑥𝑏+ 𝛿.
(1)
The rule is written this way, rather than as “move half the difference”, because it must stay in the integers: the parity
correction is what keeps 𝛿integral when 𝑑is odd. The quadratic energy
11


## Page 12

Figure 5: The carrier and its spectrum. Left: the twelve ports in their icosahedral positions, with the thirty seams drawn and
the twenty faces shaded; ports are labelled with the chart of tbl. 1. Right: the Laplacian spectrum, with the zero eigenvalue,
the three-dimensional eigenspace at 5−
√
5, the five-dimensional eigenspace at six, the three-dimensional eigenspace at 5+
√
5,
and the normalised gap marked. Generated by scripts/make_figures.py from core/carrier.py.
𝐸(𝑥) =
1
2 ∑
𝑣∈𝑉
𝑥2
𝑣
(2)
is the load energy. It never increases under the native repair.
Proposition 3. Under eq. 1 the total load is conserved and Δ𝐸= −(𝑑2 −(|𝑑| mod 2))/4.
At |𝑑| = 1 the decrease is exactly zero. The coin selects either a wait or a swap, and the swap changes the state at constant
load energy. This is the pathwise obstruction tested in sec. 7.1.
Definition 2. The load shell of total 𝑞(OPH’s shell, a level set of the total load; not to be confused with the lattice’s
coordination shells of Definition 3) is the set Σ𝑞= {𝑥∈ℤ𝑉
≥0 ∶∑𝑣𝑥𝑣= 𝑞}. The shell of total two has |Σ2| = (13
11) = 78
configurations, of which 66 are binary.
4.3
FEP: a bound on surprisal
The Free Energy Principle starts from a system with sensory states 𝑜and internal states 𝑠, and a recognition density 𝑞(𝑠).
The variational free energy is
ℱ[𝑞] = 𝔼𝑞[−log 𝑝(𝑜, 𝑠)] −𝔼𝑞[−log 𝑞(𝑠)] = 𝐷KL[𝑞(𝑠) ‖ 𝑝(𝑠∣𝑜)] −log 𝑝(𝑜),
(3)
so that minimising ℱover 𝑞at fixed 𝑜is minimising a divergence, and −ℱis a lower bound on the log evidence, the evidence
lower bound [Friston, 2010, Friston and Stephan, 2007]. The bound is tight exactly when the recognition density is the
posterior; the residual is the divergence.
In the linear-Gaussian construction, partition the carrier Laplacian 𝐿into recorded ports 𝑜and free ports 𝑥. The seam energy
is 𝐸(𝑥, 𝑜) = 1
2𝑧⊤𝐿𝑧, with 𝑧= (𝑥, 𝑜). At least one record grounds the interior block 𝐿𝑥𝑥. A record prior 𝜌(𝑜) = 1
2𝜖‖𝑜‖2 with
𝜖> 0 grounds the full precision. Integrating the resulting probability density over 𝑥gives the record surprisal [Mueller, 2026,
theorem 7]:
12


## Page 13

−ln 𝑝(𝑜) =
1
2𝑜⊤(𝑆+ 𝜖𝐼)𝑜+ 𝑑𝑜
2 ln(2𝜋) −1
2 ln det(𝑆+ 𝜖𝐼),
(4)
where 𝑑𝑜is the number of record coordinates and 𝑆is the Schur complement of 𝐿𝑥𝑥in 𝐿. The interior Gaussian determinant
appears in the raw integral of exp(−𝐸) and cancels in the probability density. With the prior fixed, eq. 4 gives an absolute
value and determines the minimum of eq. 3 over 𝑞. The probes compare this marginal route to an energy-minus-entropy
route that uses the full joint precision.
A Markov blanket is a partition of the state into interior 𝑖, blanket 𝑏and exterior 𝑒such that the interior and exterior are
conditionally independent given the blanket:
𝑝(𝑖, 𝑒∣𝑏) = 𝑝(𝑖∣𝑏) 𝑝(𝑒∣𝑏).
(5)
eq. 5 is a statement about a probability distribution. It becomes a statement about dynamics only under further conditions,
which is the content of Proposition 16.
4.4
Synergetics: tetravolume accounting
Synergetics replaces Cartesian coordinates with four signed counts whose sum is free, the quadray coordinates eq. 6. A
quadray 𝑞= (𝑎, 𝑏, 𝑐, 𝑑) is an equivalence class,
(𝑎, 𝑏, 𝑐, 𝑑) ∼(𝑎+ 𝑡, 𝑏+ 𝑡, 𝑐+ 𝑡, 𝑑+ 𝑡)
for every 𝑡∈ℝ,
(6)
so that a point is determined by differences. In the default embedding the four axes map to the vertices (1, 1, 1), (-1,
1, -1), (-1, -1, 1) and (1, -1, -1) of a regular tetrahedron, and in the integer lattice the twelve nearest neighbours of
the origin are the distinct permutations of (2, 1, 1, 0):
𝒩= { 𝜋(2, 1, 1, 0) ∶𝜋a permutation },
|𝒩| = 12.
(7)
This is the twelve-around-one configuration, the vector equilibrium of Synergetics and the kissing configuration of the A3 root
lattice, the face-centred cubic lattice [Conway and Sloane, 1999, Schütte and van der Waerden, 1953]. The neighbours are
at squared distance 8 in the default embedding, so this embedding’s unit of length is half the edge of the cube whose face
diagonals are the neighbour moves; every Euclidean number below is stated in that unit.
Definition 3. The coordination shell 𝑆𝑘is the set of lattice sites at graph distance exactly 𝑘from the origin in the graph
whose edges are the twelve neighbour moves. Halving the embedded coordinates maps the lattice onto 𝐷3 = {𝑢∈ℤ3 ∶
𝑢1 + 𝑢2 + 𝑢3 ≡0 (mod 2)}, where the graph distance has the closed form
𝑔(𝑢) = max(‖𝑢‖∞,
1
2‖𝑢‖1),
|𝑆𝑘| = 10𝑘2 + 2
(𝑘≥1).
(8)
𝑔is the gauge function of the cuboctahedron whose vertices are the twelve neighbours, so 𝑆𝑘is the set of lattice points on
the surface of the frequency-𝑘vector equilibrium.
The count 10𝑘2 +2 is the coordination sequence of the face-centred cubic lattice [Conway and Sloane, 1997] and QuadMath’s
shell formula [Friedman, 2025]; the package builds the shells by breadth-first expansion and checks the closed form 𝑔against
every site through shell 12. A coordination shell is a polyhedral surface, not a sphere: it meets several Euclidean spheres,
which is the content of Proposition 22.
The tetravolume of a lattice simplex is the absolute determinant of its three edge vectors, each written in quadray coordinates
with its fourth coordinate subtracted from the other three, divided by four, so the regular tetrahedron of the four axes has
volume one. The six solids of Synergetics’ table then have integer volumes, tetrahedron 1, cube 3, octahedron 4, rhombic
dodecahedron 6, cuboctahedron 20, truncated octahedron 96, and the package derives every entry exactly from an explicit
vertex set rather than reading it from the table (sec. 8.1). No square root appears in the accounting of the lattice’s own
solids, which is the point of the convention [Fuller and Applewhite, 1975, 1979, Edmondson, 1987]; the one solid that needs
one is the carrier’s (sec. 8.5).
Proposition 4. The quadray lattice generated by eq. 7 has 12 neighbour moves, all at squared distance 8, coordination
shells of 10𝑘2 +2 sites with cumulative counts (1, 13, 55, 147, 309) for 𝑘= 0, … , 4, and tetravolume ratios tetrahedron
13


## Page 14

1, cube 3, octahedron 4, rhombic dodecahedron 6, cuboctahedron 20, truncated octahedron 96, each derived
exactly.
4.5
One object, three vocabularies, and a dictionary this paper tests
tbl. 2 states the translation the rest of the paper tests. It is a dictionary, not an argument: each row asserts that two terms
denote the same mathematical object, and sec. 6 reports whether the assertion survives execution.
Table 2: The vocabulary dictionary this paper tests. Each row is a claim, and the last column names the registry entries
whose probes execute it; sec. 6 reports the verdicts. The last two rows are the ones where the dictionary is least conservative,
because a solid is not a graph: the agreements there are statements about the same geometry read from different constructions,
and sec. 7.10 reports where reading them as the same solid fails.
OPH
FEP
Synergetics
probed by
port (patch)
coordinate of the
generative model
quadray site
carrier-port-chart
seam
edge of the
precision graph
IVM neighbour move
gmrf-independence, kl-lattice-metri
c
load configuration
state of the
recognition
density
site occupancy
shell-counts
seam repair
variational update
local re-accounting
native-repair-law, variational-equi
valence
total load
conserved
quantity of the
dynamics
no counterpart: a load total is not a shell
frequency
native-repair-law, shell-counts
{12, 30, 20}
carrier
blanket-
structured
precision graph
icosahedron on the IVM
carrier-port-chart, gmrf-independen
ce, kissing-twelve
carrier
automorphisms
symmetries of the
precision
icosahedral group of order 120
icosahedral-symmetry
obstructed seam
cycles
frustrated cycle
space of the
precision
cycle rank 19 for the identity sheaf
sheaf-cohomology
one shell step
one divergence
step of 4 nats
one move of squared distance 8
kl-lattice-metric, tetravolume-arit
hmetic
the carrier’s
twenty faces
faces of the
simplicial complex
icosahedron of 18.5123 tetravolumes
subgraph-census, icosahedron-tetrav
olume
the
twelve-around-one
cell
neighbourhood of
one coordinate
rhombic dodecahedron of 6 tetravolumes
packing-density, dual-cells
The dictionary is deliberately conservative. It does not say that a port is a quadray site, only that the two play the same role in
the three formalisms and that the resulting numbers must therefore agree where both are defined. Where the numbers do not
agree, sec. 7 says so. The carrier’s counting invariants — 5184000 spanning trees, 19 sheaf obstructions, 120 automorphisms,
and a census of 20 triangles, 12 five-cycles, 125 matchings and 1280 Hamiltonian cycles — are derived from the same complex
in sec. 6, and each of them is a statement that two vocabularies inherit rather than assert.
14


## Page 15

5
Method: a registry whose every claim is executed, scoped and re-runnable
5.1
The unit of work
The paper’s unit is a correspondence: a claim about the same object made in two or three vocabularies, together with
executable evidence and an explicit pairwise scope.
Definition 4. A correspondence is a record 𝑐= (𝜅, 𝜌, 𝜎, 𝜋) where 𝜅names the claim in the OPH, FEP and Synergetics
vocabularies, 𝜌is the pairwise scope, one of oph-fep, fep-synergetics, synergetics-oph or three-way, naming the two
or three programmes the claim relates, 𝜎is a status from the fixed vocabulary below, and 𝜋is a probe: a total function from
no arguments to a mapping from names to values. The probe passes when every value it reports satisfies the assertion the
claim makes.
The scope 𝜌is not bookkeeping. A correspondence is a claim that two vocabularies are talking about one object, and which
two they are is part of the claim: the same icosahedral symmetry is a statement about OPH and Synergetics, while the
Markov blanket of the carrier precision is a statement about OPH and the FEP. sec. 6 reports the three pairwise ledgers
separately for that reason, and oph-fep-synergetics verify --pair synergetics-oph executes one ledger on its own.
The kind of a claim is selectable the same way: oph-fep-synergetics verify --kind disagreement runs the obstructions
alone, so a reader who wants the failures does not have to read past the identifications.
Making the probe return values rather than a boolean is the single design choice that keeps the registry honest. A boolean
can be made true by weakening a tolerance; a mapping of named values has to be right, and it can be compared against a
source that states a different value.
Table 3: The status vocabulary. It is closed: a new status requires a new definition.
Status
Meaning
exact
the probe reproduces the claim exactly: integers, rationals, or
closed-form surds
exact-on-domain
the identification holds on a stated domain and the probe asserts the
boundary
obstructed
the identification fails, and the probe exhibits the failure precisely
not-reproduced
the source states a quantity its text does not determine; recorded,
never invented
5.2
The registry, grouped by pairwise scope
tbl. 4 is the registry, reproduced from src/oph_fep_synergetics/analysis/registry.py. The grouping is the paper’s
argument in miniature: 19 claims survive as identifications, 15 fail as identifications while remaining informative, and 10
are complements in which one program supplies what another lacks without either reducing to the other. By scope, 26
claims relate OPH to the FEP, 14 relate Synergetics to OPH, 2 relate the FEP to Synergetics, and 2 entries need all three
vocabularies at once.
Table 4: The 44 correspondences, in registry order: 19 isomorphisms, 15 disagreements and 10 complementarities, scoped
oph-fep 26, synergetics-oph 14, fep-synergetics 2, three-way 2. Every row is executed by oph-fep-synergetics ve
rify.
Key
Kind
Status
Pair
What it tests
carrier-port-chart
isomorphism
exact
synergetics-oph
derived adjacency reproduces
the twenty oriented faces
spectral-gap
isomorphism
exact
oph-fep
one matrix read two ways
(mean-repair generator and
precision): 𝜆1 = 5 −
√
5,
normalised gap (5 −
√
5)/60
shell-counts
isomorphism
exact
oph-fep
the total-two load shell and its
binary sub-shell, enumerated
and checked against closed
forms — the finite domain both
oph-fep readings evaluate on
15


## Page 16

Key
Kind
Status
Pair
What it tests
native-repair-law
isomorphism
exact
oph-fep
OPH’s integer rule
reimplemented, with its
closed-form energy drop
variational-equivalence
isomorphism
exact-on-domain
oph-fep
harmonic extension,
constrained (KKT) minimiser
and perception flow coincide;
free energy at the posterior
equals the surprisal
free-energy-dictionary
isomorphism
exact-on-domain
oph-fep
Gaussian carrier prior gives
normalised surprisal, and the
residual quadratic equals the
KL between posteriors
sheaf-agreement
isomorphism
exact
oph-fep
the cellular sheaf Laplacian is
the carrier Laplacian
certificate-species
isomorphism
exact
oph-fep
the certificate species of the
shell and its spectral gap
tree-number
isomorphism
exact
oph-fep
one determinant, two readings:
5184000 counts spanning trees
and normalises the FEP’s
Gaussian interface integral
gmrf-independence
isomorphism
exact
oph-fep
the seam graph is the
conditional-independence graph
of the precision: dependence
exactly on the seams; blanket 5,
variance 0.1944444
sheaf-cohomology
isomorphism
exact
oph-fep
patch agreement is the
global-section condition:
dim 𝐻0 = 1, dim 𝐻1 = 19 on
the graph; with the faces, Betti
numbers (1, 0, 1)
icosahedral-symmetry
isomorphism
exact
synergetics-oph
order 120, transitive on ports,
seams and faces: 60 rotations,
60 reflections
eigenspace-irreducibility
isomorphism
exact
synergetics-oph
the eigenspaces are the group’s
non-trivial real irreducibles:
dimensions (3, 3, 5) forced by
the group and established by
the commutant test
law-equivariance
isomorphism
exact
synergetics-oph
every transition commutes with
every automorphism; 4 orbits,
3 outcome multisets
kl-lattice-metric
isomorphism
exact
fep-synergetics
one IVM step costs four nats
and squared norm 8𝑘2 costs 4𝑘2
pr01-obstruction
disagreement
obstructed
oph-fep
ungated repair does not
terminate; the gate can give
different normal forms
free-energy-name-collision
disagreement
obstructed
oph-fep
“free energy” names
inequivalent functionals
mediator-obstruction
disagreement
obstructed
oph-fep
pairwise boundary reduction
needs a threshold
stationary-blanket
disagreement
obstructed
oph-fep
a stationary blanket is not an
attractor
solenoidal-freedom
disagreement
obstructed
oph-fep
stationary blankets leave
entropy production free
d-stability
disagreement
obstructed
oph-fep
stability is not generic in the
coupling
schedule-path
disagreement
obstructed
oph-fep
iteration schedules are not
interchangeable
reduction-criterion
disagreement
obstructed
oph-fep
competing reduction criteria
disagree
16


## Page 17

Key
Kind
Status
Pair
What it tests
existence-without-reachability
disagreement
obstructed
oph-fep
existence of an equilibrium
does not imply reachability
directed-gradient
disagreement
obstructed
oph-fep
direct and adjoint gradients
differ
four-cycle-local-minimum
disagreement
obstructed
oph-fep
the four-cycle has spurious
local minima
cuboctahedron-vs-icosahedron
disagreement
obstructed
synergetics-oph
twelve vertices is not one solid:
tetravolume 20 against
18.5123 …
shell-vs-free-energy
disagreement
obstructed
fep-synergetics
the coordination shell (a
polyhedral norm) and the
divergence (a Euclidean one)
determine neither each other;
levels (36, 52) are shared
quadray-ivm
complementarity
exact
synergetics-oph
IVM coordinates supply exact
geometry; the tetravolume
table is derived from vertex sets
equilibrium-propagation
complementarity
exact
oph-fep
the energy-based learning
identity holds exactly
replicator-potential
complementarity
exact
oph-fep
the replicator field is a
Shahshahani gradient for
symmetric payoffs, and not
otherwise
kissing-twelve
complementarity
exact
synergetics-oph
both twelve-around-one
arrangements attain the kissing
number: the cuboctahedral
rigidly, the icosahedral with
slack
icosahedron-tetravolume
complementarity
exact
synergetics-oph
the carrier’s solid is 18.5123 …,
squared exactly 175 + 75
√
5
packing-density
complementarity
exact
synergetics-oph
Voronoi cell 14/12/24,
tetravolume six, density
𝜋/(3
√
2)
tetravolume-arithmetic
complementarity
exact
three-way
integer nats, a surd gap, and an
integral table beside a
non-integral carrier
effective-resistance
complementarity
exact
oph-fep
effective resistance 11/30, 7/15,
1/2 by distance class, Kirchhoff
index 28 twice over
two-record-resistance
isomorphism
exact
oph-fep
the inverse effective resistance
sets the Schur energy of a
pair’s record contrast
vertex-shells
disagreement
obstructed
synergetics-oph
shells (5, 5, 1) against the
lattice’s (12, 42, 92);
intersection array (5, 2, 1; 1, 2,
5)
subgraph-census
isomorphism
exact
synergetics-oph
20 triangles, 12 pentagons, 125
matchings, 1280 Hamiltonian
cycles, counted by exhaustion
angle-defect
isomorphism
exact
synergetics-oph
OPH’s derived chart is the
icosahedron’s triangulation, so
the angular-defect accounting
applies unchanged: 𝜋/3 at each
port, 4𝜋in total, which is 2𝜋𝜒
twelve-around-one-families
disagreement
obstructed
synergetics-oph
the phrase covers solids with
5184000 and 331776 = 5762
spanning trees
dual-cells
complementarity
exact
synergetics-oph
the cell and the cuboctahedron,
14/12/24 against 12/14/24,
volumes 6 and 20
17


## Page 18

Figure 6: The registry as a scope matrix. Each row is one correspondence in registry order; the columns are its pairwise
scope, its kind, its status, and whether the entry carries an executable probe, with cells coloured by category. The figure is
drawn from analysis/registry.py, so an entry cannot appear in the matrix without appearing in verify.
Key
Kind
Status
Pair
What it tests
conjugation-norm
complementarity
exact
three-way
conjugation fixes the integer
counts and moves the volume,
whose norm is 50
repair-is-conditioning
isomorphism
exact-on-domain
oph-fep
on the binary shell each native
repair is the uniform reference
conditioned on its visible
datum and minimises a free
energy over the fibre; off the
shell it fails on 60 seams
The registry is the source of truth for the paper. The prose in sec. 6, sec. 7 and sec. 8 says what the probes found; it does not
add claims the probes do not carry. fig. 6 shows the same 44 rows as a matrix of scope, kind and status, so that the shape
of the registry is visible without reading the table.
5.3
Reading a source without copying it
The three source programs are read at pinned commits, and the pins are data, not prose: data/sources.json records
the commit, the branch, the licence terms and, for each borrowed identifier, the file, the line range and the symbol. The
command oph-fep-synergetics sources --vault-root <root> re-reads the live checkouts and reports match, moved,
absent or unavailable for each pin.
Three rules follow from the licences, and they are enforced rather than recommended.
OPH is dual-licensed: its software, including the Lean development and the model code, is Apache-2.0, while its papers,
book, documentation, figures, data and ledgers are CC BY-NC-SA 4.0 [FloatingPragma, 2026].
This package therefore
reimplements the software behaviour with attribution and reads no data or ledger from OPH. The carrier adjacency is the
18


## Page 19

clearest case: it is derived from OPH’s twenty oriented faces and then checked against them, rather than transcribed from
any table, and no OPH data file is read at run time.
FEP_Formal is CC BY 4.0 and is cited by DOI and by topic identifier [Friedman, 2026b]. The catalogue is read through its
machine-readable metadata (config/catalogue_metadata.yaml), which is the schema-2 roster sealed at fep-001 through
fep-168; the generated projection config/topics.yaml is never authored or edited, because FEP_Formal declares it a
projection and regenerates it.
QuadMath is Apache-2.0 [Friedman, 2025]. The constants this package needs are reimplemented and the source is cited
by symbol: the neighbour moves, the embedding, the shell formula and the tetravolume convention. This paper quotes no
QuadMath prose.
Two corrections in the provenance record are worth reporting, because both were caught by checking rather than by care.
An earlier draft of this package attributed the carrier table to a Lean file that does not contain it; the attribution was false
and the table was replaced by a derivation. And an earlier draft cited a QuadMath commit that had been superseded; the
live checkout had moved, and the pin was updated. Both are recorded in docs/PROVENANCE.md. A pin is a point-in-time
reading, and the discipline that matters is that the reading is re-done rather than remembered.
5.4
Why exactness, not agreement
It would be easier to report that three literatures “converge”. The registry does not do that, for a reason that sec. 7 makes
concrete: two programmes can agree on a number while disagreeing on the object the number is about. Existence without
reachability Proposition 20 is the clearest example. Both programmes can be right that an equilibrium exists, and both can
be right about its value, while one of them is wrong to infer that a trajectory reaches it. Agreement on the value would have
hidden the disagreement about the dynamics.
The registry therefore distinguishes what is computed from what is claimed about it, and the statuses are assigned to claims.
This is why the paper reports more disagreements than a synthesis-oriented reading would, and why the disagreements carry
the same weight as the agreements.
19


## Page 20

Figure 7: The registry as a matrix. Each row is one correspondence, coloured by kind (isomorphism, disagreement, comple-
mentarity) and marked by status (exact, exact-on-domain, obstructed). The figure is drawn from the same registry object
that oph-fep-synergetics verify executes, so it cannot show an entry the probes do not carry. Generated by scripts/m
ake_figures.py from analysis/registry.py.
6
Agreement: the identifications each pairwise scope survives
The registry records 19 isomorphisms across the three pairwise scopes. This section states each one, gives the exact values
the probe reports, and says what the identification does not cover. fig. 7 shows the whole registry at a glance, and fig. 8
shows the same registry split by the pair of programmes each claim relates.
6.1
The port chart is a consequence, not a transcription
Proposition 1 is exact in the strongest sense available: the derivation returns 12 ports, 30 seams and 20 faces; every degree
is 5; the twenty triangles of the derived graph are exactly the twenty oriented faces; and the derived adjacency equals the
declared adjacency. The probe reports all six facts separately, so a failure of any one of them is visible rather than masked
by the conjunction.
This matters for a reason that is methodological rather than mathematical. A published table of adjacencies can be copied
with a transposition and still produce a plausible graph. Deriving the graph from the faces, and then checking that its
triangles reproduce the faces, closes the loop: a transposition anywhere in the source data would break the round trip. The
same pattern is used wherever this package reads a table that has consequences.
6.2
The spectrum is closed-form
Proposition 2 gives the full spectrum.
The probe reports the eigenvalues as 2.763932 three times, six five times, and
5 +
√
5 ≈7.236068 three times, with the zero eigenvalue once, and checks that the Laplacian really is 5𝐼−𝐴rather than a
normalised variant.
The gap is the quantity that carries dynamical meaning. Normalising by the total degree,
𝛾=
𝜆1
2|𝐸| = 5 −
√
5
60
= 0.04606553,
(9)
gives a number that is invariant under rescaling and is the one that appears in the relaxation estimates below. The corre-
sponding repair chain is the symmetric stochastic matrix
20


## Page 21

Figure 8: The three pairwise ledgers. Each triangle is one scope, named by the two programmes it relates; inside each
triangle are that scope’s counts by kind and its statuses, computed from the registry the command line executes. The panels
are drawn from analysis/registry.py by scripts/make_figures.py, so a scope that gains or loses a claim changes its
triangle.
𝑃= 𝐼−
𝐿
2|𝐸|,
(10)
whose rows sum to one, whose entries are non-negative because every degree is at most 2|𝐸|, and whose second eigenvalue
is 1 −𝛾= 0.9539345 — the certificate-species correspondence, reported by the probe as a row-stochastic, symmetric,
non-negative matrix with exactly that second eigenvalue and induced infinity norm 1. The chain of eq. 10 is the natural
object behind the phrase “the shell mixes at the spectral gap”: its mixing is governed by 𝛾, and 𝛾is a property of the carrier
alone.
The chain is also what the integer rule does on average, and the probe checks that on the rule itself. Averaging eq. 1 over every
seam and both coins, from each of the 78 configurations, returns 𝑥−𝐿𝑥/(2|𝐸|) in exact rationals: the mean current across
a seam is half its mismatch for either parity. OPH states the same identity as one gradient step of the seam-disagreement
energy under its uniform calibration [FloatingPragma, 2026], so eq. 10 is derived from the law and the Laplacian is the law’s
own mean generator.
6.3
The shell is counted exactly
Definition 2 gives the count. The probe reports |Σ2| = (13
11) = 78 total-two configurations and 66 binary ones, and checks the
enumeration against the closed forms (13
11) (stars and bars: two units of load over twelve ports) and (12
2 ) (two occupied ports
out of twelve), which share no code with it. The check is worth having because the enumerated set is the one that the repair
census below partitions.
6.4
The native repair law is reproduced, not approximated
The native-repair-law correspondence reimplements OPH’s rule on exact integers and checks it over all 4680 transitions
of the 78-state shell. The rule itself is eq. 1; the closed form for the energy drop is Δ𝐸= −(𝑑2 −(|𝑑| mod 2))/4, reported by
the probe as a string so that the formula, and not merely a sample of its values, is part of the record.
The reimplementation is exact in the sense that matters: it is not a numerical approximation of OPH’s rule but the same
integer arithmetic, and the energy identity is checked rather than assumed. What the correspondence does not say is that
OPH’s rule is the right dynamics; that is an empirical question, and OPH’s own status document records that no retained
laboratory comparison selects a native record law [FloatingPragma, 2026].
21


## Page 22

Proposition 5. Let 𝐿be the carrier Laplacian partitioned into recorded ports 𝑜and free ports 𝑥, equip 𝑜with the prior in
eq. 4, and let 𝑥∗(𝑜) = −𝐿−1
𝑥𝑥𝐿𝑥𝑜𝑜be the harmonic extension of the records. The posterior has mean 𝑥∗(𝑜) and precision 𝐿𝑥𝑥.
At that posterior, the variational free energy takes the value ℱG(𝑜) = −ln 𝑝(𝑜) [Mueller, 2026, theorem 7], with
𝑆= 𝐿𝑜𝑜−𝐿𝑜𝑥𝐿−1
𝑥𝑥𝐿𝑥𝑜,
ℱG(𝑜) −ℱG(𝑜0) =
1
2(𝑜⊤𝑆𝑜−𝑜⊤
0 𝑆𝑜0) + 𝜌(𝑜) −𝜌(𝑜0).
(11)
The probe uses two antipodal carrier records with prior precision 0.5 and repeats the checks on a four-node chain. It computes
the settled interior of eq. 11 three ways that share no linear solve. The harmonic extension solves the interior block. The
constrained minimiser solves one saddle-point (KKT) system for min 1
2𝑧⊤𝐿𝑧subject to 𝑧𝑜= 𝑜over all coordinates and a
multiplier per record. The prior contributes a fixed term at fixed records, so the same constrained solution minimises the
relative entropy of a mean shift at fixed joint precision. This is the constrained relative-entropy reading of OPH’s premise
A3. And the predictive-coding flow perceive reaches the same point in 143 iterations with residual 9.169439e-11, from every
start tried. The probe also evaluates the free energy at the posterior directly, as energy minus entropy, and finds it equal to
the surprisal: the divergence in eq. 3 has numerical residual 0 because the bound is tight at the posterior.
The boundary control clamps a single record on an oriented seam with an asymmetric spring.
The energy is piecewise
quadratic, with twice the stiffness for positive disagreement. Its unique minimiser is at zero disagreement, while numerical
integration and a half-Gaussian calculation give a negative posterior mean. The equality of minimiser and posterior mean in
Proposition 5 depends on the Gaussian assumption.
6.5
Each native repair is exact conditioning
The identifications above read one object in two vocabularies. This one relates the native law itself to a free-energy functional,
and it is exact [Mueller, 2026, Mueller et al., 2026].
A repair of a seam conserves the load on the seam and touches no other port. The seam’s total load together with every
load off the seam is therefore the same before and after the repair. That pair is the datum the repair leaves visible, and the
configurations that share it are its fibre.
Proposition 6. Let 𝑢be the uniform law on the 66 binary configurations and 𝐾𝑒the law of one native repair of the seam
𝑒under its fair coin. For every binary configuration 𝑥and every seam 𝑒, 𝐾𝑒(𝑥, ⋅) is 𝑢conditioned on the datum 𝑜of 𝑥at 𝑒.
With 𝜋(𝑠, 𝑜) = 𝑢(𝑠) for 𝑠in the fibre of 𝑜, 𝐾𝑒(𝑥, ⋅) is the unique minimiser of
ℱ[𝑞] = 𝔼𝑞[ln 𝑞(𝑠) −ln 𝜋(𝑠, 𝑜)]
(12)
over the laws 𝑞on the fibre, and the minimum is the surprisal −ln 𝜋(𝑜) of the datum.
The probe compares the two sides of Proposition 6 in all 3960 cases, both orientations of every seam from every binary
configuration, by routes that share no code: the integer rule on one side, the reference and the datum on the other. A
seam that holds one unit has a fibre of two configurations and the repair resamples the unit’s side with equal weights, which
happens in 1200 cases; every other seam has a fibre of one. The surprisal is log 33 in the first case and log 66 in the second.
Evaluated on eq. 12, a coin with any bias pays exactly its divergence from the fair coin above that minimum, and a rule that
always waits or always swaps pays log 2.
Because 𝐾𝑒is a conditional expectation of 𝑢, every law 𝑝on the shell obeys 𝐷(𝑝‖ 𝑢) = 𝐷(𝑝‖ 𝑝𝐾𝑒) + 𝐷(𝑝𝐾𝑒‖ 𝑢). Each repair
lowers the relative entropy to the reference by exactly 𝐷(𝑝‖ 𝑝𝐾𝑒), which the probe checks on a mixed state for every seam.
OPH states both facts for the binary shell [FloatingPragma, 2026, Mueller et al., 2026].
The status is exact-on-domain, and the probe asserts the boundary. With the uniform law on all 78 configurations as
reference the identity fails at the doubled configurations, on the 60 seams that meet a doubled port out of 360 off-shell cases:
the repair moves one unit across with certainty, and conditioning would spread the weight over three configurations.
In the FEP’s vocabulary each native repair is a perception step whose recognition density is the exact posterior, with the
uniform reference as prior and the repaired datum as observation, so the bound of eq. 3 is tight at every step. The functional
each step minimises is eq. 12. It is a different object from the load energy of eq. 2, which sec. 7.1 shows to be constant along
the binary shell while every step there is exact inference. Registered as repair-is-conditioning.
22


## Page 23

Figure 9: Two panels read from the executed probes. Left: the sandwich interval with its lower bound 0.4318644 and upper
bound 3.125, and the Bayesian surprise 0.875 marked inside it. Right: three quantities at the same records: minimum
interface energy 0.25, unnormalised record potential -1.555415, and normalised record surprisal 1.714912. Generated by scr
ipts/make_figures.py from core/free_energy.py and core/controls.py.
6.6
The free-energy dictionary has a precise scope
Three quantities are routinely called “surprisal”, and the source literature distinguishes five: (numerical, bayesian, surprisal,
critic, meaningful) [Mueller, 2026]. In the linear-Gaussian model this package computes three of them, and tbl. 5 reports
what happens. The registry entry is free-energy-dictionary.
Table 5: Three surprises in the linear-Gaussian carrier probe. The residual norm and Bayesian surprise describe a change
of records. Record surprisal uses the new records alone. The sandwich bound brackets the Bayesian value [Mueller, 2026,
theorem 7 and table 7].
Quantity
Definition
Value in the probe model
numerical surprise
residual norm ‖𝑔‖ at the old rest
2.5
Bayesian surprise
1
2𝑔⊤𝐿−1
𝑥𝑥𝑔
0.875
record surprisal
−ln 𝑝(𝑜) under the proper joint
model
3.994912
variational free energy
eq. 3 at the posterior
3.994912
The probe reports a sandwich with lower bound 0.4318644 and upper bound 3.125, and the Bayesian value 0.875 lies inside
it. The energy drop at the new records and the divergence between posteriors give the same value by independent routes.
The variational free energy equals record surprisal at the posterior and exceeds it by the divergence from the posterior away
from that optimum.
The asymmetric seam gives a check outside this domain. Ground one port and read the other with a symmetric record prior.
Records of opposite sign and equal magnitude have different surprisals, because the prior and the normaliser cancel in their
difference and the seam energies do not agree. The centred quadratic record formula cannot represent this difference.
6.7
A sheaf computes the same Laplacian
The scalar carrier construction instantiates the cellular sheaf formulation of agreement [Mueller, 2026, sec. 7.2 and theorem
7].
Proposition 7. Give every port the stalk ℝand every seam the identity restriction. The resulting cellular sheaf on 𝐾has
30 edges, Laplacian equal to the carrier Laplacian, one-dimensional space of global sections (the constants), and injective
record maps at every port.
23


## Page 24

The probe reports all of this, and it reports two negative controls: a chain sheaf with one record has a one-dimensional space
of global sections, and a chain sheaf with no record has none. The controls matter because a sheaf whose global sections are
one-dimensional for every choice of records would not be saying anything about the carrier. Registered as sheaf-agreement.
The sheaf formulation is the one that generalises. A scalar identity sheaf is the simplest case; the same construction with
non-identity restrictions is how the FEP’s precision structure would enter, and the harmonic-extension identity of Proposition
5 is the same computation written in sheaf language [Hansen and Ghrist, 2019, Robinson, 2014]. The agreement reported
here is therefore not a coincidence of two implementations but a statement that the carrier’s Laplacian is the coboundary
Laplacian of a sheaf, which is the bridge [Friedman, 2026c] proposed.
6.8
The carrier counts its own spanning trees
Proposition 8. Delete any row and the matching column of the carrier Laplacian. The determinant of the remaining 11×11
matrix is an integer, and by Kirchhoff’s matrix-tree theorem [Kirchhoff, 1847] that integer is the number of spanning trees
of the seam graph:
𝜏(𝐾) = 5184000 =
1
12
12
∏
𝑖=2
𝜆𝑖.
(13)
The probe reports the determinant computed in exact rational arithmetic, and, separately, the product of the eleven nonzero
eigenvalues computed in floating point: the first is 5184000, the second agrees with it to nine digits. The two computations
are not redundant. The eigenvalues include the surds 5 ±
√
5 three times each, so an exact integer arising from a product of
irrationals is a statement about the arithmetic and not about the sampling; the floating-point product is reported next to it
because the disagreement between the two is the honest error bar on the spectral route.
The count is not decoration.
One determinant, two readings: in OPH’s the integer counts the seam sets that connect
every port without a cycle; read as a Gaussian precision with one port pinned (the gauge that makes it proper), the same
determinant normalises the linear-Gaussian interface integral the FEP dictionary factorises through. Either way the carrier’s
spanning trees are a finite, exactly known family, so any claim about the existence of an agreement configuration can be
phrased as a claim about a member of a set of known size. sec. 7.7 is the obstruction that shows why the size of that set
does not by itself settle reachability.
6.9
The Gaussian reading makes the seam graph the independence graph
Use 𝐿+𝐼/𝑛as the precision of a proper Gaussian field on the 𝑛carrier ports. The positive diagonal shift grounds the constant
mode and leaves the off-diagonal entries unchanged. The precision’s zeros determine conditional independence [Dempster,
1972, Rue and Held, 2005]: two ports are independent given the rest exactly when their precision entry vanishes, which here
happens exactly when they are not joined by a seam, so conditional dependence lives exactly on the seams. The probe reports
the independence pattern as a 12 × 12 matrix over all 60 ordered pairs of distinct ports, checks that it equals the adjacency
pattern, and checks that the shifted precision is positive definite.
Every seam’s pairwise precision is -1. A port’s Markov blanket consists of its 5 neighbours. For the separate intrinsic field
centred to sum to zero, the covariance is 𝐿+ and each port’s variance is 0.1944444, exactly 7/36. Vertex transitivity gives this
value by dividing the reciprocal spectrum sum by 𝑛. The probe computes the sum over the rationals as tr((𝐿+ 𝐽/𝑛)−1) −1,
where 𝐽is the all-ones matrix.
Conditioning the proper Gaussian on all other ports gives the graph result. For the centred intrinsic field the same conditioning
fixes the sum of the remaining pair, including for nonseams. sec. 7.3 separately tests whether a blanket in a stationary density
constrains the dynamics.
6.10
Two cohomology dimensions, one global section
Proposition 9. Give every port the stalk ℝand every seam the identity restriction, and let 𝛿0 be the resulting coboundary
on the seam graph. Then dim 𝐻0 = 1 (the constants), dim 𝐻1 = 19, and 𝛿0 has rank 11, so the Euler characteristic of the
graph is -18. Adding the twenty faces as two-cells, with 𝛿1 the signed face-seam incidence, 𝛿1 has rank 19: the surface has
Betti numbers (1, 0, 1) and Euler characteristic 2, and OPH’s oriented faces sum to a fundamental cycle.
The probe reports both coboundaries’ shapes and ranks in exact arithmetic, all the cohomology dimensions, both Euler
characteristics, the check that 𝛿1𝛿0 = 0, and the dimensions of the same quantities for a chain sheaf used as a control.
24


## Page 25

The two numbers have concrete readings, which is the reason for reporting them separately rather than as one identity.
The global sections are the assignments of equal values to adjacent ports, so dim 𝐻0 = 1 says that the only way for
every seam to agree is for every port to carry the same value: the agreement is a constant. The cohomology dimension
dim 𝐻1 = 19 measures seam assignments modulo coboundaries of port assignments. Its value is the cycle rank of the seam
graph, |𝐸| −|𝑉| + 1 = 30 −12 + 1. Every disagreement that a port assignment actually produces lies in the image of 𝛿0.
Both dimensions are forced by the graph [Hansen and Ghrist, 2019, Ghrist, 2014].
The second half of Proposition 9 reconciles two numbers the paper uses. The graph’s Euler characteristic is -18; the surface’s,
which the angle-defect count of Proposition 13 reaches through Gauss-Bonnet, is 2. The difference is the faces: the face-seam
incidence has rank 19, equal to the cycle rank, so the face boundaries span the graph’s cycle space and no first cohomology
survives on the sphere. The oriented faces sum to a cycle – every seam is traversed once in each direction. This checks the
coherence of the declared orientation used by the rotation and reflection count Proposition 10.
6.11
The symmetry group closes the census
Proposition 10. The automorphism group of the carrier has order 120, is closed under composition, and acts transitively
on the 12 ports, the 30 seams and the 20 faces. It splits into 60 rotations and 60 reflections, and the Laplacian eigenspaces
of dimensions (3, 3, 5) are irreducible under it. Across all group elements and all 78 total-two configurations, the multiset of
descent outcomes takes 3 values, on 4 orbits of configurations.
The probe reports each part separately: the group order from an exhaustive backtracking search, closure under composition,
the orbit decomposition of ports, seams and faces, the split into rotations and reflections, the orbit ranks and commutant
dimensions of the three nontrivial eigenspaces against their dimensions, and the equivariance of the descent law over the full
product of 120 group elements and 78 configurations.
Three readings follow, and they are the reason this correspondence is in the registry rather than in an appendix. First,
the symmetry is a symmetry of the object, so any claim about the carrier that is not group-invariant is a claim about a
labelling. Second, the eigenspaces being irreducible means the multiplicities (3, 3, 5) are forced: they are the dimensions
of the icosahedral group’s representations, so no perturbation that keeps the symmetry can change the multiplicity pattern.
The eigenvalues are not forced: the permutation representation is multiplicity-free, so every invariant symmetric matrix has
these eigenspaces, each with its own value. Third, and most useful for this paper, the repair law is equivariant transition
by transition: for every group element, configuration, oriented seam and coin, repairing the relabelled configuration on the
relabelled seam gives the relabelled result, so the multiset of outcomes is preserved, and the probe finds no witness that
would break it. The census numbers in sec. 7.1 are therefore invariants, and the 4 orbits are the right unit of analysis. The
registry carries this section as three entries: icosahedral-symmetry, eigenspace-irreducibility and law-equivarianc
e.
The rotation and reflection counts require the orientation of the faces, and that detail is a correction worth reporting. Testing
orientation by the set of directed edges of a face fails on a closed surface, because such a set already contains every edge in
both directions; the test used here compares the cyclic order of each face, rotated so that its smallest port leads, and only
then does it separate 60 rotations from 60 reflections.
6.12
Information cost is the synergetic metric over two
Proposition 11. For two Gaussians of equal covariance Σ centred at the origin and at a lattice site 𝑧, the Kullback–Leibler
divergence is half the squared Mahalanobis norm of the separation; with Σ = 𝐼in the default embedding,
𝐷KL(𝒩(0, 𝐼) ‖ 𝒩(𝑧, 𝐼)) =
1
2‖𝑧‖2,
(14)
and the IVM’s first neighbour shell sits at squared norm 8, so one step costs 4 nats and the sites 𝑘𝑚along a neighbour move
𝑚cost 4, 16, 36 and 64 nats for 𝑘= 1, … , 4, that is 4𝑘2.
The probe reports the divergence for each of the twelve neighbour moves, the squared distance of a move, and the costs of
the sites 𝑘𝑚for the first four multiples [Kullback and Leibler, 1951].
This is the shortest of the identifications and the one with the widest reach, because it is the only place in the registry where
the FEP’s own currency and Synergetics’ own metric are the same function. The lattice supplies the distances; the divergence
turns them into costs; and because every squared distance between lattice sites is a multiple of eight in this embedding, every
cost is an integer number of nats, a multiple of four. The number four is a statement about the unit of length, not about the
lattice: rescaling the covariance rescales every cost by the same factor, and with the covariance equal to the squared prime
25


## Page 26

vector one step costs half a nat. What the identification fixes is the ratio of costs, which is the ratio of squared distances. It
also fixes a convention that the paper uses everywhere else: a “one-step cost” is a divergence, not an energy, and the two
differ by the factor of two that eq. 14 makes explicit.
6.13
The carrier’s subgraphs are the solid’s own counts
Proposition 12. Counting the seam graph’s subgraphs by exhaustion returns 20 triangles, 12 induced (chordless) five-
cycles, 125 perfect matchings and 1280 Hamiltonian cycles counted as undirected edge sets, and the first of those is exactly
the number of faces the port chart was derived from.
The probe enumerates rather than quotes: every three-vertex subset for the triangles, every five-vertex subset for the
pentagons, and a memoized recursion for the matchings and the Hamiltonian cycles, all from the adjacency matrix alone.
The triangle count is where two derivations meet. The carrier’s faces were derived from OPH’s oriented base faces in sec. 6.1,
and the census then rediscovers 20 triangles in the graph that derivation produced. A count that agreed by construction
would prove nothing; this one is computed from the adjacency and checked against the face list afterwards.
The induced five-cycles are the links of the twelve ports, one around each, and so they are the twelve pentagonal faces of the
dual dodecahedron, which is the first place in the paper where the dodecahedron appears without being named. The perfect
matchings and the Hamiltonian cycles have no counterpart in any of the three programmes, and that is the reason to report
them: they are properties of the object all three are about, available to anyone who can read its adjacency.
6.14
Curvature is exact in units of pi
Proposition 13.
Five equilateral triangles meet at every port, so each port carries an angle defect of 1/3 pi and the
twelve ports carry 4 pi in total, which is 2𝜋𝜒for a triangulated sphere: Descartes’ theorem on the total angular defect, the
combinatorial form of Gauss-Bonnet, read off the face list [Federico, 1982].
The probe computes each defect as twice pi minus the incident faces times pi over three and returns it as an exact rational
multiple of pi rather than as a decimal: 1/3 at each port, 4 in total, against the surface Euler characteristic 2 that the
cohomology of Proposition 9 reaches from the other direction. OPH’s derived chart is the icosahedron’s triangulation, so the
angular-defect accounting Synergetics supplies for that solid applies unchanged: the vertex figure the accounting measures
is the very five-face figure every port carries. The identity is the reason twelve five-fold vertices close into a sphere rather
than a dish, and it is exact for the same reason the tree number is: rational quantities over a finite complex. Registered as
angle-defect.
fig. 10 draws this section and the two that follow it: the resistance ladder, the shells around a port, the census (Proposition
12), the defect (Proposition 13), and the dual pair.
6.15
What the agreements do not say
One identification relates OPH’s dynamics to a free-energy functional: on the binary shell each native repair minimises eq. 12
(Proposition 6). It holds step by step on that shell. It leaves the load energy of eq. 2 what it was, and it says nothing off
the shell. Several of the other identifications are statements about counting, and most of the rest are statements about the
carrier, its spectrum, its repair rule as an integer map, the Gaussian free energy on it, and the sheaf whose Laplacian it is.
Every one of those is a statement about a finite object that both programmes can be given.
The remaining identifications are structural rather than numerical, and the paper reports them with the same care: a
symmetry group of order 120 is exact arithmetic, a spanning-tree number is an exact integer, and a cohomology dimension
is an exact rank. What none of them supplies is a claim that either programme’s learning dynamics is the other’s.
The next section reports the places where an identification fails, and most of them fail for one reason: the authors of each
programme attach a dynamical claim to that shared object, and the dynamical claims do not agree.
26


## Page 27

Figure 10: The carrier read as a network. Panel one gives the exact resistance ladder between port pairs with the Kirchhoff
index; panel two colours the twelve ports by their graph distance from one port and annotates the intersection array; panel
three counts the subgraphs; panel four compares the twelve-around-one cell with its dual, whose counts and tetravolumes are
read from the geometry package. Every number is computed from the executed package at draw time rather than transcribed.
27


## Page 28

Figure 11: The descent census and a mismatch-one excursion. Left: the 4680 transitions of the shell, partitioned into 120 strict
descents, 600 zero descents and 3960 no-ops. Right: a one-unit excursion around the carrier through distinct configurations
while the quadratic energy stays at 0.5, so the trajectory has no descent to offer and no unique limit. Generated by script
s/make_figures.py from core/controls.py::gated_descent_census.
7
Disagreement: the obstructions, each stated exactly
An obstruction is a result. Each of the 15 claims below was proposed as an identification, and each fails in a way that can
be stated exactly. The section is ordered by how much the failure costs a synthesis: the first distinguishes confluence from
termination for OPH’s native rule, the next two block the identification of “free energy”, the remainder block dynamical
identifications that the literature makes routinely, and the last two block identifications that are made by counting rather
than by dynamics.
7.1
The native law does not terminate
OPH’s PR-01 distinguishes its native seam rule from the integer and transactional contracts that carry its consensus theorems
[FloatingPragma, 2026]. In the Lean library, demoCarrier_not_confluent and demoCarrier_terminates concern a separate
Boolean copy-repair example. The native rule instead fails termination on the carrier’s total-load-two shell. At mismatch one
a native attempt waits or swaps the odd unit; the swap changes the state at constant load energy [Mueller, 2026, sec. 6.2].
Registered as pr01-obstruction.
Proposition 14. Over the 78 configurations of Σ2, the 30 seams and the two coins, the native rule has 4680 transitions.
Of these, 120 strictly lower eq. 2, 600 leave it unchanged, and 3960 leave the configuration unchanged. In particular the
zero-descent transitions are not rare: they outnumber the strict descents, and the multiplicities of the native transitions take
exactly the values (1, 2, 50, 52).
The mechanism is eq. 1 itself, and the energy identity is Proposition 3. At |𝑑| = 1 the parity correction absorbs the whole
of the move: coin zero’s current is 1 and coin one’s is 0, so one coin performs a swap at constant energy and the other does
nothing at all. The probe reports both currents and both energy decreases, which are 0 and 0. fig. 11 shows the census and
an excursion.
The same draws fix the two probabilities the source states for the configuration with one unit on each of two adjacent ports
[Mueller, 2026]. One attempt is a uniform seam and a fair coin. The configuration stays where it is with probability 13/15
and moves to each of 8 neighbouring configurations with probability 1/60, and every draw leaves eq. 2 unchanged. Two
transitions with the same energy decrease differ in surprisal, so no coeﬀicient converts the path surprisal of the native law
into its energy decrease. OPH’s model states the kernel and both values [FloatingPragma, 2026]; the probe enumerates the
draws and checks the result against a count of the seams that meet exactly one of the two occupied ports.
Proposition 15. Count only state-changing repairs as rewrite steps. On the whole shell, the ungated relation has 0 normal
forms and admits a cycle. The binary configurations form one closed, strongly connected class reachable from every state,
28


## Page 29

so any two descendants of the same state can reach a common successor: the relation is confluent but does not terminate.
Under the broken-seam gate, every step strictly lowers load energy. The gated relation terminates, with the 66 binary states
as normal forms. Each doubled state reaches 5 distinct normal forms, so the gated relation is not confluent.
The state graph in Proposition 15 concerns individual paths. Under a uniform seam and fair coin, the binary-shell attempt
kernel includes the waits. Its transition-count matrix is symmetric with all row and column totals equal to the number of
draws. The uniform law is stationary, connectivity gives its uniqueness, and the computed spectral gap equals the port gap
in Proposition 2. Each single-seam kernel is a conditional expectation (Proposition 6); the mixture also preserves the uniform
law. Data processing makes relative entropy to that law nonincreasing for every initial binary-shell distribution. The probe
separately checks a trajectory from a point mass and the spectral bound on chi-square divergence. This bound does not
assign a universal decay rate to relative entropy. Thus a path can keep its load energy constant while the ensemble loses
relative entropy.
7.2
Three distinct values at fixed records
On the carrier the minimum interface energy min𝑥𝐸(𝑥, 𝑜) is 0.25. The unnormalised record potential −ln ∫exp(−𝐸(𝑥, 𝑜)) 𝑑𝑥
is -1.555415. Their difference -1.805415 equals the interior Gaussian term 1
2 ln det 𝐿𝑥𝑥−𝑑𝑥
2 ln(2𝜋), computed independently
as -1.805415.
For the normalised model, record surprisal and variational free energy at the posterior both equal 1.714912. This value
includes the record prior energy and the record normalisation term 1.432412 from eq. 4. The interior term cancels against the
corresponding term in the joint normaliser [Mueller, 2026, theorem 7]. Two chain controls with identical Schur complements
and the same prior give identical surprises despite different interior determinants and unnormalised potentials. Registered
as free-energy-name-collision.
7.3
A stationary blanket does not bound the dynamics
This control follows Mueller’s proposition using the three-state blanket example of Da Costa et al. [Mueller, 2026, proposition
11; Da Costa et al., 2021].
Proposition 16. Order the state as exterior 𝜂, blanket 𝑏and interior 𝜇, let the precision be 𝐻= (
2 1 0
1 2 1
0 1 2 ), so that 𝐻𝜇𝜂= 0,
and let 𝑄be antisymmetric with 𝑄𝜇𝜂= 𝑞. The diffusion d𝑧= −(𝐼−𝑄)𝐻𝑧d𝑡+
√
2 d𝑊has stationary law 𝒩(0, 𝐻−1) for
every 𝑞, which satisfies eq. 5 exactly. Its interior drift iṡ 𝜇= 2𝑞𝜂+ (𝑞−1) 𝑏−2𝜇, which at 𝑞= 0.75 has coeﬀicients (1.5,
-0.25, -2) on (𝜂, 𝑏, 𝜇), so two states with equal blanket and interior values move differently whenever their exteriors differ,
with cross-drift difference 1.5.
The stationary law factorises; the dynamics does not. The conditional independence of eq. 5 is a property of a distribution,
and a system can have it at stationarity while every trajectory through the state space mixes the interior and the exterior.
The probe reports the factorisation and the coupling in the same mapping, so the reader cannot take the first without also
seeing the second.
This is the disagreement that the critical literature presses [Bruineberg et al., 2022, Biehl et al., 2021, Aguilera et al., 2022],
and the probe says it precisely: the blanket is real as a statistical statement and insuﬀicient as a dynamical one. A synthesis
that needs a boundary to do something — to shield, to persist, to constitute an agent — needs an additional condition on
the drift, and that condition is not implied by eq. 5.
7.4
Stationary blankets leave entropy production free
Mueller’s entropy-production control uses a family of diffusions with a common stationary law, following work on nonreversible
Gaussian diffusions [Mueller, 2026, proposition 9; Hwang et al., 1993; Lelièvre et al., 2013].
Proposition 17. For d𝑥= −(Γ−𝑄)𝐻𝑥d𝑡+
√
2Γ d𝑊with Γ, 𝐻positive definite and 𝑄antisymmetric, the law 𝒩(0, 𝐻−1) is
stationary for every 𝑄and the entropy production rate is tr(𝑄⊤Γ−1𝑄𝐻). In two dimensions with Γ = 𝐻= 𝐼and 𝑄= 𝑞( 0 1
−1 0 )
the drift has eigenvalues −1 ± 𝑖𝑞and the entropy production is 2𝑞2: (0, 0.5, 4.5) at 𝑞∈(0, 0.5, 1.5). Stationarity therefore
does not fix the entropy production, and cannot be used to infer it.
The solenoidal component of the drift — the part that circulates without changing the stationary density — is invisible to
the stationary law and fully visible in the entropy production. fig. 12 draws both panels.
This is the sharpest of the disagreements, because it is a statement about what a stationary-blanket argument can and cannot
conclude. The dissipative structure is constrained; the circulating structure is not [Ao, 2004, Seifert, 2012, Godrèche and
29


## Page 30

Figure 12: Stationary blankets and their freedom. Left: the three drift components at 𝑞= 0.75, showing that the drift couples
the exterior and the interior even though the stationary law factorises. Right: entropy production against the solenoidal
parameter 𝑞, which is 2𝑞2 and therefore free, while every member of the family is stationary. Generated by scripts/make_f
igures.py from core/controls.py::blanket_dynamics_check and solenoidal_check.
Luck, 2019]. A free energy argument that works only from stationarity therefore cannot reach the thermodynamic quantities
that the solenoidal part controls. Registered as solenoidal-freedom.
7.5
Stability is not generic in the coupling
Definition 5. A real matrix 𝐴is stable when every eigenvalue has negative real part, and D-stable when 𝐷𝐴is stable for
every positive diagonal matrix 𝐷[Arrow and McManus, 1958]. A positive diagonal 𝐷is a choice of local repair gains that
leaves the rest point 𝐷𝐴𝑥= 0 where it was.
Proposition 18. Let 𝐴= (
−1 2
0
2 −1 3
0 −3 −1 ) be Mueller’s stable three-node control matrix [Mueller, 2026, sec. 4.1 and proposition
10], and let 𝐷𝑘= diag(𝑘, 1, 1). The family 𝐷𝑘𝐴has characteristic polynomial 𝑧3 + (𝑘+ 2)𝑧2 + (10 −2𝑘)𝑧+ 6𝑘, and its Routh
first column is
(1,
𝑘+ 2,
2 (10 −𝑘2)
𝑘+ 2
,
6𝑘),
(15)
so for 𝑘> 0 every root has negative real part exactly when 𝑘<
√
10 ≈3.162278. At 𝑘= 4 the column is (1, 6, -2, 24),
whose two sign changes certify 2 right-half-plane roots. 𝐴is stable and not D-stable.
The eigenvalues at 𝑘= 4 are -6.288814 and 0.144407 with imaginary part ±1.948191: the complex pair has positive real part,
and the system spirals out.
What is lost is a free-energy form. A linear law admits the Helmholtz decomposition 𝐴= −(Γ −𝑄)𝐻with Γ, 𝐻positive
definite and 𝑄antisymmetric – the Ornstein-Uhlenbeck form whose stationary density is ∝exp(−1
2𝑥⊤𝐻𝑥) – exactly when
𝐴is stable: with Γ = 𝐼the Lyapunov equation 𝐴𝑃+ 𝑃𝐴⊤= −2𝐼gives 𝐻= 𝑃−1 and 𝑄= 1
2(𝐴𝑃−𝑃𝐴⊤) [Khalil, 2002]. For
the unscaled 𝐴the source’s certificate is exact: 𝑃is positive definite with det 𝑃= 4.259259, the residual of the Lyapunov
equation is 2.531698e-15, and 𝑄is antisymmetric [Mueller, 2026]. At 𝑘= 4 no positive-definite solution exists, because 𝐷4𝐴is
unstable, so one positive local gain removes the form without moving the rest. Mueller embeds the control in a nonlinear field
[Mueller, 2026, proposition 10]. With 𝑊= 4(𝐴+𝐼) and a logistic activation, the bounded neural field 𝑓(𝑣) = −𝑣+𝑊𝜎(𝑣)+𝑏
with 𝑏= −𝑊( 1
2, 1
2, 1
2)⊤has 𝑓(0) = 0 and Jacobian exactly 𝐴at the origin, which the probe confirms with zero error. The
obstruction is therefore not that the algebra fails. It is that stability is a property of a point in a parameter space, and the
identification of a repair rule with a stable descent needs D-stability (Definition 5), every admissible gain on the stable side.
Here the stable region is 𝑘<
√
10, which is not the whole admissible range. Registered as d-stability.
30


## Page 31

Figure 13: D-stability. Left: the eigenvalue locus of 𝐷𝑘𝐴in the complex plane for 𝑘∈(0.5, 1, 2, 4, 7), with the imaginary axis
drawn; the pair that crosses into the right half plane is marked. Right: the third Routh entry 2(10 −𝑘2)/(𝑘+ 2) of eq. 15,
which changes sign at 𝑘=
√
10 and is -2 at 𝑘= 4. Generated by scripts/make_figures.py from core/controls.py::d_s
tability_eigenvalues and routh_first_column.
Figure 14: The mediator threshold. Left: the pairwise margin 3𝜀−1 against 𝜀, -0.4 at 𝜀= 0.2 and crossing zero at the
threshold 0.3333333. Right: the reduction error at and below the threshold, showing that the worst-case error floor is 1
while retaining the record is exact. Generated by scripts/make_figures.py from core/controls.py::mediator_pairwi
se_margin and reduction_control.
7.6
Reduction needs a threshold
The mediator model and the trace proof of the threshold come from Mueller [Mueller, 2026, proposition 8]. The probe here
also constructs and checks the rank-one blocks at the threshold.
Proposition 19. For the three-block Gaussian model with mediator coupling 𝜀, a pairwise representation of the reduced
boundary exists exactly when 𝜀≥0.3333333. Below the threshold the pairwise margin is 3𝜀−1, which is -0.4 at 𝜀= 0.2; at
the threshold it is 0 and the decomposition exists, is positive semidefinite, and reproduces the target.
The threshold is not an artefact of the parametrisation. Necessity is a trace inequality: each pairwise block must carry the
off-diagonal entry ± 1
3 of the target, so a positive semidefinite block has diagonal sum at least 2
3, and the three blocks together
need at least two units of trace against the target’s 1 + 3𝜀. Suﬀiciency is an explicit decomposition, and at 𝜀= 0.3333333
every block in it is rank one: the pairwise form sits on the boundary of the semidefinite cone there, and below it the cone is
left. Two further disagreements belong with it. In Mueller’s reduction control [Mueller, 2026, proposition 4], the two criteria
disagree: the same observation scores 1, the first criterion scores 1 afterwards and the second -1, the response distance is 2,
and the worst-case error has a lower bound of 1 — so the choice of criterion changes the verdict rather than the precision.
And the iteration schedule is not a detail: the centred contrast after twelve steps differs between the full and the half-damped
schedules by exactly 0.02304085 in absolute value, so a claim about “the” schedule path has to name the schedule [Mueller,
31


## Page 32

Figure 15: The absolute difference between the full and the half-damped schedules’ centered contrasts, one exact value per
step count from one to twelve. The schedules share every answer while the contrast difference decays to 0.02304085 after
twelve steps: agreement and dissipation are state functions, and a finite schedule is not. Generated by scripts/make_figu
res.py from core/controls.py.
2026]. Registered as mediator-obstruction, with reduction-criterion and schedule-path beside it.
OPH’s executable observer runs the reduction test on the native law itself [FloatingPragma, 2026, Mueller, 2026]. The source
is uniform over the binary configurations. The observer repairs one seam, records the signed current, and chooses its second
seam from that record alone. Over 264 equally weighted paths the final occupation of port zero is 35/264. With the selector
disconnected, or driven by an independent record with the same seam frequency, it is 1/6, the occupation of the uniform
binary state. The reduction-criterion probe enumerates the paths and checks the result against a case analysis that
enumerates nothing. A summary that drops the record keeps the action frequencies and loses the response.
7.7
Existence does not imply reachability
The clipped map and the relaxation orbit reproduce Mueller’s two-node control [Mueller, 2026, proposition 15].
Proposition 20.
The two-node relaxation has Jacobian eigenvalues 1 ± 2 𝑖, a relaxation orbit whose radius stays in
[0.8400708, 1.053391] with period 7.575, and a synchronous clipped map whose orbit minimum radius is 1 — a different
object with a different bound.
Mueller gives the bound 0.84 and period 7.58 [Mueller, 2026, proposition 15]. The probe measures 0.8400708 and 7.575. Both
numbers are reported, because the difference between a source’s bound and a package’s measurement is exactly the kind of
thing a reader needs to see rather than to trust. fig. 16 draws the orbit.
The disagreement is subtle and worth spelling out. Both programmes can prove that a fixed point exists, and both can
compute it; the probe finds the complete fixed-point set of the clipped map by exact case analysis over its nine clipping
regimes, and it is the origin alone. The synchronous orbit’s radius bound is measured after a transient, which the probe
records. The relaxation dynamics has a Jacobian with positive real part at the fixed point, so the fixed point is unstable, and
the orbit is bounded away from it. Existence and reachability are different questions, and a synthesis that infers the second
from the first is wrong in a way that no amount of algebraic care will fix. Registered as existence-without-reachability.
7.8
Direct and adjoint gradients differ
Mueller gives a two-dimensional example [Mueller, 2026, appendix B]. At its stable fixed point, with the non-symmetric
Jacobian 𝐽and loss derivative 𝑐, a forward nudge measures the direct response −𝐽−1𝑐= (1, 0), while the gradient of the
loss with respect to the parameters needs the adjoint −𝐽−⊤𝑐= (1, 2), the recurrent back-propagation of [Pineda, 1987].
32


## Page 33

Figure 16: Existence without reachability. Left: the relaxation orbit in the plane, with the annulus [0.8400708, 1.053391]
that contains it; the orbit neither escapes nor reaches the origin. Right: the radius against time, showing the period 7.575
and the contrast with the synchronous clipped map, whose orbit is confined to a different annulus. Generated by scripts/
make_figures.py from core/controls.py::two_node_relaxation_orbit.
Only the adjoint is the gradient; the two coincide exactly when 𝐽is symmetric, which is the case an energy-based network
supplies. A learning rule that reads the forward response therefore does not follow the gradient of a directed network, and
the disagreement is not a numerical artefact. Registered as directed-gradient.
7.9
The four-cycle has an incompatible local minimum
In Mueller’s four-cycle control [Mueller, 2026, sec. 4.1], each edge carries an equality constraint. The compatible states are
all-zeros and all-ones, with no violations, while the state 0011 violates two edges. Every single flip of that state also leaves
two violations, so it is a (non-strict) local minimum of the violation count sitting on a plateau, and a descent method that
only accepts strictly improving moves cannot leave it. The probe reports the violations after each of the four flips, all of
which are 2.
This is the mechanism behind the traps that repair-as-descent arguments have to exclude by hand. It is also a concrete
instance of the general point: a descent guarantee is a statement about a landscape, and a landscape with plateaux has local
minima that are not global ones. Registered as four-cycle-local-minimum.
7.10
Twelve vertices is not one solid
The IVM’s first shell and the carrier’s vertex set both contain twelve points. That count is not an identification, and the
probe exhibits the difference rather than the coincidence.
Proposition 21. The twelve nearest neighbours, joined when their squared distance is eight, form the cuboctahedron graph:
4-regular, 24 edges, eight triangles and six squares, with tetravolume 20. The carrier’s graph is 5-regular with 30 edges and
twenty triangles, and its solid has tetravolume 18.5123. The two graphs have the same vertex count and are not isomorphic.
The probe reports both degree sequences, both edge counts, both face decompositions, both volumes, and the fact that the
icosahedral realisation of the carrier is isomorphic to the carrier and not to the neighbour graph. The tetravolumes settle
it without reference to the graph at all: 20 against 18.5123 …, one integral and one algebraic. The failure costs more than
a footnote, because the two solids support opposite dynamical readings. The cuboctahedron’s neighbours are arranged so
that each vertex sees four others, which gives a lattice with a four-fold local order; the carrier’s ports see five, which is what
puts the surd
√
5 into the spectral gap (5 −
√
5)/60 eq. 9 and what the FEP’s blanket count of 5 inherits. A synthesis that
identified the two by their common twelve would import a coordination number the carrier does not have.
33


## Page 34

Both solids are kissing arrangements of twelve spheres around one, which is where the confusion starts, and they attain the
kissing number in different ways. With the twelve centres at twice the sphere radius from the centre, the cuboctahedron’s
squared edge over squared circumradius is 1, so every sphere touches four of its neighbours as well as the centre: the rigid
arrangement of the densest packing. The icosahedron’s ratio is 2−2
√
5/5 ≈1.105573, so its spheres touch the centre and none
of each other. The slack is what Fuller’s jitterbug uses to carry the vector equilibrium continuously through the icosahedron
to the octahedron [Fuller and Applewhite, 1975, Verheyen, 1989], and the volume is not conserved along the way. Synergetics
therefore does not conflate the two solids – it distinguishes them and connects them by a motion – and the obstruction is to
a reader who identifies them by their count.
The same coincidence is why this entry is a disagreement rather than a complementarity: the neighbour cell of sec. 8.6 is the
rhombic dodecahedron, of tetravolume six, whose twelve faces face the twelve neighbours and whose graph has six four-valent
and eight three-valent vertices. Three distinct solids in this geometry have twelve-fold features, and only one of them is the
carrier. Registered as cuboctahedron-vs-icosahedron.
7.11
The shell index and the free energy determine neither each other
Proposition 22. Let 𝐹(𝑧) = 1
2‖𝑧‖2 be the equal-covariance divergence of Proposition 11 at a lattice site 𝑧, and 𝑘(𝑧) its
coordination-shell index (Definition 3). Neither is a function of the other. The first five shells carry (1, 3, 4, 7, 7) distinct
values of 𝐹, and the levels (36, 52) are each carried by two shells. On shell 𝑘the divergence ranges from 2𝑘2 (from 2𝑘2 + 2
when 𝑘is odd) up to 4𝑘2, and reaches 4𝑘2 only at the twelve vertices 𝑘𝑚.
The probe walks the shells, collects the distinct values of 𝐹on each, reports the counts per shell, the levels that appear on
more than one shell, and the extreme values of 𝐹on every shell through 12. A shell is not a level set of 𝐹, and a level of 𝐹
is not confined to one shell. The lower bound does give one inequality: 𝐹≤𝐿forces 𝑘≤√𝐿/2, so a bounded free energy
bounds the shell index from above. What fails is the identification, in both directions.
The reason is exact and easy to state once the shell has its closed form. The shell index is the gauge 𝑔(𝑢) = max(‖𝑢‖∞, 1
2‖𝑢‖1)
of eq. 8, a polyhedral norm whose unit ball is the vector equilibrium, while 𝐹is a Euclidean quadratic form whose level sets
are spheres. A sphere and a cuboctahedron touch only along the cuboctahedron’s twelve vertex directions, which is why 𝐹
attains its shell maximum only at 𝑘𝑚and why the face points of shell 𝑘sit closer to the centre than its vertices, by up to a
factor of
√
2 in distance.
fig. 17 draws both halves of the statement.
This is the paper’s only coordinate-level obstruction. It does not say that the FEP’s functional is wrong or that Synergetics’
metric is wrong; it says that a composite vocabulary that treats “shell” and “free energy” as two names for the same
stratification will be inconsistent at the first shell with more than one radius, which is shell two. Registered as shell-vs-fr
ee-energy.
7.12
The shells around a port are not the lattice’s shells
Proposition 23. The ports at graph distance one, two and three from a port number (5, 5, 1), the graph is distance-regular
with intersection array (5, 2, 1; 1, 2, 5) and diameter 3, and the lattice’s own shells at the same three radii number (12, 42,
92).
The two sequences (5, 5, 1) and (12, 42, 92) disagree from the first term: the carrier’s first shell is a port’s 5 neighbours
because the icosahedron is five-valent, while the IVM’s first shell holds 12 sites because the lattice is twelve-coordinated by
construction. The proposal that a port’s neighbourhood is a lattice shell therefore fails on the smallest possible test, and no
choice of radius repairs it: the carrier is finite with diameter 3, so its shells stop, while the lattice’s do not.
What survives is a structural statement rather than an identification. The carrier is distance-regular, with intersection array
(5, 2, 1; 1, 2, 5) read off a single representative per shell, so its neighbourhood structure is rigid: from any port, at any distance,
the outward and backward degrees are the same [Brouwer et al., 1989]. That is a stronger fact than the identification it
replaces, and it follows from the symmetry: the group of order 120 is transitive on the pairs of ports at each distance, and a
distance-transitive graph is distance-regular.
7.13
Twelve around one names a family, not an object
Proposition 24. The phrase covers at least two solids with different counting invariants. The icosahedron is five-valent
with 5184000 spanning trees; the cuboctahedron is four-valent, carries 8 triangles and 6 squares, 32 perfect matchings, and
331776 spanning trees, which is 576 squared.
34


## Page 35

Figure 17: The coordination shell against the free energy. Left: the points of the face-centred cubic lattice 𝐷3 in the plane
𝑢3 = 0, coloured by the closed-form shell index 𝑔of eq. 8; in this slice the shells are nested squares, and the overlaid circles
are levels of the free energy, which cross several shells. Right: for every shell through 12, the range of 𝐹over the shell as
a bar and each distinct level as a dot, with the levels shared between shells highlighted and the curves 2𝑘2 and 4𝑘2 drawn;
the maximum is attained only at the twelve vertices of each shell. A bounded free energy bounds the shell, and neither is a
function of the other. Generated by scripts/make_figures.py from geometry/lattice.py.
35


## Page 36

Figure 18: Two bar series against the radius from a port, on a logarithmic count axis. The carrier’s ports number (5, 5, 1) at
graph distances one, two and three, while the lattice’s shells hold (12, 42, 92) sites at the same radii, so the sequences disagree
at the first shell and no radius repairs the reading. The carrier is distance-regular, and the title annotates its intersection
array (5, 2, 1; 1, 2, 5). Generated by scripts/make_figures.py from core/network.py and geometry/lattice.py.
This is the same failure sec. 7.10 reports, one level down. There the shared quantity was the vertex count; here it is the phrase
itself. A reader who is told that a structure is “twelve around one” has been told the coordination number and nothing else,
and the coordination number does not determine the object: the two solids differ in degree, in triangle count, in matching
count, and in the spanning-tree number, whose ratio is exactly 125/8.
The counting invariant is what does the work. Spanning trees are computed by the same determinant (eq. 13; Proposition 8) in
both cases, on adjacency matrices derived independently – the carrier’s from OPH’s oriented base faces, the cuboctahedron’s
from the neighbouring lattice sites – so the numbers are comparable even though the objects are not the same. What the
registry records is that a claim of the form “the carrier is the twelve-around-one solid” is under-determined, and that the
remedy is a refinement rather than a tolerance. Registered as twelve-around-one-families.
7.14
What the obstructions have in common
The obstructions fall into three kinds. The first is a plateau: zero-descent moves (PR-01), plateaux in the violation landscape
(the four-cycle), and constant-energy swaps all describe dynamics that cannot be ordered into a descent. The second is a
missing invariant: the entropy production of a stationary blanket, the stability of a coupling outside its stable region, and
the direction of a gradient all vary within a class that a coarser vocabulary treats as fixed.
The third is the one a correspondence registry is best placed to catch: an invariant too coarse to identify the object it is asked
to identify. Twelve vertices do not determine a solid (Proposition 21), the phrase “twelve around one” covers two solids it
does not separate (Proposition 24), a port’s neighbourhood is not a lattice shell (Proposition 23), and a shell index does not
determine a free energy (Proposition 22). None of these failed because a number disagreed; each failed because the shared
quantity was never an identification, and a table of matched counts would have reported each as a success.
A synthesis that wants to identify the three programmes must therefore add structure that excludes plateaux, fixes the
missing invariants, and refines the coarse ones. That is a research programme, not a corollary, and the next section reports
what each programme can contribute to it.
36


## Page 37

Figure 19: Two panels that separate the solids the phrase twelve-around-one can name. Left: the cuboctahedron’s degree,
edges, triangles and squares beside the carrier’s — 4 against 5 neighbours, 8 triangles and 6 squares against twenty triangles
and none, with the two tetravolumes annotated. Right: the spanning trees the same determinant counts — 5184000 for the
carrier against 331776 for the cuboctahedron, which is 576 squared — on a logarithmic axis. Generated by scripts/make_
figures.py from core/invariants.py and geometry/lattice.py.
8
Complementarity: structures one programme supplies and another needs
The 10 complementarities are not identifications and not failures. Each is a case where one programme supplies a structure
that another needs and does not have, without either reducing to the other. They are the constructive part of the paper:
most of them are geometric, the rest are a learning identity, a potential flow and a resistance metric, and two need all three
vocabularies at once.
8.1
Synergetics supplies the accounting in which the carrier’s geometry is exact
The carrier of Definition 1 is a graph. To draw it, to count its shells, and to state the volume of the polyhedra it bounds,
a graph is not enough: a metric is needed, and the choice of metric decides which numbers come out exact. Synergetics
supplies one in which they all do.
Proposition 25. The twelve quadray neighbour moves of eq. 7 are distinct, all have squared norm 8, and their minimum pair-
wise squared distance is also 8, so the twelve form a kissing configuration of radius
√
2. The coordination shells, enumerated,
have (1, 12, 42, 92, 162) sites for 𝑘= 0, … , 4 and cumulative counts (1, 13, 55, 147, 309). The tetravolume table t
etrahedron 1, cube 3, octahedron 4, rhombic dodecahedron 6, cuboctahedron 20, truncated octahedron 96 is
derived exactly: each solid is the convex hull of an explicit integer vertex set built from the lattice – the tetrahedron from
the four axes, the cube from the axes and their negatives, the octahedron from the pairwise axis sums, the cuboctahedron
from the twelve neighbours, the rhombic dodecahedron as their Voronoi cell, and the truncated octahedron by cutting the
frequency-three octahedron at the thirds of its edges – and its volume, computed in rational arithmetic, is divided by the
unit tetrahedron’s.
The complementarity is exact and one-directional. OPH’s dynamics does not need quadray coordinates, and the FEP does
not need a lattice. But the moment the carrier is drawn, or its shells are counted as geometry rather than as graph distance,
the quadray convention is what makes the statements exact rather than approximate. The derivation matters because a table
is a claim: the cube’s three and the truncated octahedron’s ninety-six are conventions about which cube and which truncated
octahedron (the cube whose face diagonals are the prime vector, and the truncated octahedron whose edge is), and only a
vertex set fixes that choice. fig. 20 shows the shells and the volume ratios.
The kissing number is the point at which the two programmes touch.
The carrier has 12 ports of degree 5; the IVM
has 12 neighbours at squared distance 8; and the geometric icosahedron’s adjacency is isomorphic to the derived carrier
adjacency, which the probe checks by constructing a witness. The coincidence is not a numerological one: twelve is the
kissing number in three dimensions [Schütte and van der Waerden, 1953, Conway and Sloane, 1999], and both arrangements
37


## Page 38

Figure 20: The IVM accounting. Left: shell sizes 10k^2 + 2 for 𝑘= 0, … , 8 with the cumulative count overlaid; the first shell
is the twelve-around-one configuration. Right: tetravolume ratios on a logarithmic scale, with the unit tetrahedron at one,
so that the cube, octahedron, rhombic dodecahedron, cuboctahedron and truncated octahedron are all integers. Generated
by scripts/make_figures.py from geometry/lattice.py.
attain it. They attain it differently, which the kissing-twelve probe measures exactly: the cuboctahedral arrangement is
rigid, its spheres touching their neighbours, while the icosahedral one has slack, its squared edge over squared circumradius
being 2 −2
√
5/5 ≈1.105573 rather than 1 (sec. 7.10). The carrier is the loose arrangement; the lattice is the tight one.
8.2
The same quadratic sequence appears three times
Fuller’s geodesic subdivision and Goldberg’s dual construction produce the same pair of quadratic sequences, and the IVM
shell count is one of them.
Proposition 26. For every frequency 𝑓≥1: the geodesic subdivision of the icosahedron has 10𝑓2 + 2 vertices, 30𝑓2 edges
and 20𝑓2 faces; its Goldberg dual has 20𝑓2 vertices, 30𝑓2 edges and 10𝑓2 + 2 faces. Hence the vertex count of one is the face
count of the other, and both are the IVM shell count of eq. 8.
This is the clearest complementarity in the paper, and the least mysterious. Duality swaps vertices and faces, so the two
counts trade places, and Euler’s formula with twelve five-valent vertices fixes the rest. The IVM shell count is the same
quadratic for a combinatorial reason: the frequency-𝑓shell of the vector equilibrium has eight triangular and six square
faces of frequency 𝑓, and cutting each square along a diagonal turns it into the frequency-𝑓geodesic icosahedron on the
same vertex set, which is the move Fuller’s jitterbug makes continuously [Fuller and Applewhite, 1975, Verheyen, 1989]. The
geodesic counts are the 𝑇= 𝑓2 case of the Caspar-Klug classification of icosahedral shells [Caspar and Klug, 1962, Coxeter,
1973]. What Synergetics adds is the insistence that the volume accounting is exact in the same convention, which is what
makes the subdivision usable as a discretisation rather than only as a drawing.
8.3
Energy-based learning supplies the gradient the repair law lacks
Equilibrium propagation is the cleanest case in which an energy is not merely a bound but a computational device: perturb
the output, let the system relax, and read the gradient from the difference in the relaxed energy [Scellier and Bengio, 2017].
It belongs to energy-based learning rather than to the FEP, which reaches the same territory through predictive coding:
inference by settling a precision-weighted prediction-error energy approximates back-propagation [Whittington and Bogacz,
2017, Millidge et al., 2022]. The probe is a scalar energy model, chosen so that every quantity has a closed form.
Proposition 27.
At perturbation 𝛽, the closed-form contrast is -1.2499987500012502 and the relaxed contrast is
-1.24999874998033, against an exact gradient of -1.25. The closed-form error is 2.092015e-11 and the distance to the gradient
is 1.249999e-06, matching the predicted 𝑂(𝛽) scaling.
The complementarity is that an energy provides the objective and the relaxation dynamics provides the implementation, and
38


## Page 39

Figure 21: Three appearances of one sequence. The geodesic vertex count 10f^2 + 2, the Goldberg face count, and the IVM
shell count are the same function of the frequency, while the geodesic face count and the Goldberg vertex count are 20f^2.
The plot shows both sequences against 𝑓and annotates the duality. Generated by scripts/make_figures.py from geomet
ry/subdivision.py.
the identity between them is exact to the stated order in the perturbation. It needs a symmetric interaction, which is why
the directed case of sec. 7.8 is an obstruction. Neither OPH’s repair rule nor Synergetics’ accounting gives such an identity,
and neither needs one.
8.4
Selection supplies a potential flow, for symmetric payoffs
Proposition 28. For a symmetric payoff 𝐴, the replicator fielḋ 𝑥𝑖= 𝑥𝑖((𝐴𝑥)𝑖−𝑥⊤𝐴𝑥) equals 𝑀(𝑥)∇𝐹with 𝐹= 1
2𝑥⊤𝐴𝑥and
mobility 𝑀= diag(𝑥) −𝑥𝑥⊤, at every point checked, with the direct potential rate 0.1344 agreeing with the rate computed
through the mobility and the field norm 0.2119623 at the reported point. For a non-symmetric payoff the same identity fails
at every point checked.
The replicator equation for a symmetric (partnership) game is a gradient flow in the Shahshahani metric [Shahshahani, 1979,
Taylor and Jonker, 1978, Hofbauer and Sigmund, 1998], and the mean payoff is a Lyapunov function for it. The probe checks
the identity, not merely the sign, and checks its boundary: with an asymmetric payoff the field is still 𝑀(𝑥)𝐴𝑥, but no
potential reproduces it.
The relevance to the three programmes runs through inference. The discrete replicator equation is Bayes’ rule, with types as
hypotheses and fitness as likelihood [Harper, 2009], so the FEP’s posterior updating and selection share one dynamical form,
and this is the one place where a potential is unambiguously a Lyapunov function of a dynamics that appears in population
biology rather than in neuroscience. It supplies a template: a synthesis that wants a free-energy descent for OPH’s repairs
should look for the metric in which the repair field is a gradient, and the replicator case shows what such a metric looks like
when it exists. Registered as replicator-potential.
8.5
The carrier’s own solid has an exact algebraic volume
The tetravolume table of sec. 8.1 is a table of integers, and the carrier’s own solid is not in it. It has an exact volume all the
same, and the fact that the volume is exact while being algebraic rather than integral is the sharpest statement of what the
synergetic convention buys.
Proposition 29. The regular icosahedron whose edge is the prime vector has tetravolume 18.5123, equal to 5
2(3
√
2 +
√
10)
exactly, and the square of that tetravolume is 342.7051, which is exactly 175 + 75
√
5. Its twelve vertices are 5-valent, and its
39


## Page 40

twenty triangles are all congruent.
The probe computes the volume twice. The exact route places the icosahedron of edge two at (0, ±1, ±𝜑) and its cyclic
permutations, with 𝜑= (1 +
√
5)/2, finds its twenty facets by an exact supporting-plane test, and sums signed tetrahedra in
the field ℚ(
√
5); dividing by the regular tetrahedron of the same edge gives the tetravolume as
√
2 times 15/2 + 5/2 sqrt
5 exactly, with no rounding anywhere. The floating-point route applies the divergence theorem to the measured surface and
divides by the same tetrahedron, and the two agree to the last digit, so the agreement is an identity rather than a tolerance
chosen to make it hold.
Two details of the computation are worth recording because they are corrections rather than conveniences. The first is that
a face list is an unordered set of triangles: ten of the twenty faces enumerated from the vertex set come out facing inward
(10 of them), and the signed sum cancels exactly before they are reoriented. The second is that the tetravolume convention
is a convention: changing the unit tetrahedron rescales every entry in the table, and the claim made here is the ratio the
convention defines and not an absolute volume.
Why this is a complementarity rather than an identification is now visible. OPH supplies the graph, and the graph’s faces
are triangles; Synergetics supplies the convention in which those triangles have exactly computable volumes; and neither
programme as written states the volume of the solid its own graph bounds. The table’s integral entries and the carrier’s
algebraic entry belong to the same accounting, which is what makes the disagreement of sec. 7.10 a statement about solids
and not about conventions.
8.6
The densest packing in three dimensions has the twelve-around-one cell
Proposition 30. The Voronoi cell of the twelve neighbours has 14 vertices, 12 rhombic faces and 24 edges, with Euler
characteristic two and no orientation flips. Its tetravolume is 6, and the ball of radius half the prime vector inscribed in it
has tetravolume 4.442883, so the packing fraction is 0.7404805, equal to 𝜋/(3
√
2).
This is the paper’s clearest case of a number that can be derived instead of read. The tetravolume of the cell is computed
from the halved squared norm at each of its fourteen vertices, using the same quadratic form that makes the neighbour
distances exact, and it arrives at six: the same six the synergetic table lists for the rhombic dodecahedron. A reader who
suspects that a table entry is a stylised fact can therefore check it by an independent route, which is the pattern this package
uses wherever a source table has consequences.
The density is then derived rather than quoted, and it is exact. The cell has integer vertices, so its volume is computed in
rational arithmetic and its tetravolume is exactly 6; the inscribed ball has radius
√
2, half the prime vector, so its tetravolume
is exactly 𝜋
√
2; and the density is their ratio, 𝜋
√
2/6 = 𝜋/(3
√
2) = 0.7404805. That number is the Kepler density of the
face-centred cubic arrangement, of which the IVM’s twelve neighbours are the local picture. It is optimal among lattice
packings [Gauss, 1840] and among all packings [Hales, 2005, Hales et al., 2017]. The complementarity is that Synergetics’
twelve-around-one is not a drawing convention but the local configuration of the densest packing in three dimensions.
The carrier does not inherit that optimality, and the same cell is the reason the disagreement of sec. 7.10 is not a quibble. The
densest packing’s kissing arrangement is the rigid cuboctahedral one, whose cell is the rhombic dodecahedron of tetravolume
six; the carrier’s solid is the loose icosahedral arrangement, of tetravolume 18.5123, with a 5-regular graph. Both put twelve
spheres around one, and they are different objects.
8.7
Three number systems and no common unit
Proposition 31. The three programmes’ exact quantities live in different number systems. The FEP’s divergence for one
IVM step is the integer 4 nats (with unit covariance in the default embedding); the carrier’s normalised spectral gap, a
relaxation rate of OPH’s mean repair operator, is the quadratic surd (5 −
√
5)/60 ≈0.04606553; and Synergetics’ table is
integral, with the twelve-neighbour cell at 6, while the carrier’s own solid lies in
√
2 ℚ(
√
5).
This is one of the two entries in the registry that need all three vocabularies at once – the other is the conjugation norm of
sec. 8.10 – and both are complementarities rather than identifications for that reason. The three quantities are attached to
different objects (a mean shift, a relaxation, a solid), so none is the “cost of a step” in another’s terms, and no common unit
converts one into another: an integer, a surd and an element of a biquadratic field cannot be made equal by rescaling. A
reader who wants one number has to choose a currency and state the exchange rate; a reader who wants the object has three
exact readings of it.
40


## Page 41

8.8
The precision also carries a resistance metric
Proposition 32. Ground the Laplacian at one port and solve 𝐿𝑥= 𝑒𝑖−𝑒𝑗over the rationals [Klein and Randić, 1993]: the
effective resistance across a seam is 11/30, at distance two 7/15, and across the carrier 1/2, over (30, 30, 6) pairs in the three
classes, summing to a Kirchhoff index of 28.
In the Gaussian reading, this resistance is the variance of the intrinsic potential difference. The probe two-record-resis
tance eliminates all free ports and compares the resulting pairwise Schur complement with the grounded network solve for
every pair. For records 𝑜𝑖, 𝑜𝑗, the Schur term in eq. 4 is (𝑜𝑖−𝑜𝑗)2/(2𝑅𝑖𝑗). The proper prior then contributes its own energy
and the record normaliser.
Across the carrier, the three classes contribute 28 in total, and the same total is twelve times the reciprocal spectrum sum
7/3 from sec. 6.9, which is a cross-check between a rational system solved pair by pair and a closed-form spectral identity.
The agreement is exact and the two routes share no arithmetic. Registered as effective-resistance.
8.9
The twelve-around-one cell and the cuboctahedron are duals
Proposition 33. The cell of the twelve neighbours and the cuboctahedron are dual edge for edge: 14 vertices against 12,
12 faces against 14, and 24 edges each, with Euler characteristic two each and tetravolumes 6 and 20, the cuboctahedron’s
to the cell’s in the ratio 10/3.
The dual pair is derived twice, and the two derivations meet. The cell is built from the constraints of the twelve lattice sites,
and the cuboctahedron from the neighbouring sites themselves; the cuboctahedron is also the cube’s line graph, which the
probe asserts by an explicit isomorphism search over the twelve vertices rather than by comparing degrees. sec. 8.6 uses the
cell for a density; this claim records that its partner is the other solid Synergetics names for the same arrangement, and that
the two accountings are related by exactly 10/3.
8.10
Conjugation fixes the counts and moves the volumes
Proposition 34. Conjugating
√
5 while fixing
√
2 leaves the counting invariants alone and moves the geometric ones. The
spanning-tree number 5184000 is a rational integer and is fixed; the icosahedron’s tetravolume 18.5123,
√
2 times 15/2 + 5
/2 sqrt 5, maps to 2.700908, and the two multiply to exactly 50, computed in field arithmetic.
This is the sharpest statement the paper can make about the difference between its two kinds of claim, and it is why sec. 9
separates them. A count over a finite complex is an integer, so the Galois symmetry of the quadratic field cannot see it; a
volume in the same field is not an integer, and the symmetry moves it to a value that no solid of the table has. The product
of the two is an integer, 50, so the conjugation is a check rather than a curiosity: a transcription error in either volume would
break a product that is exactly rational. Registered as conjugation-norm.
8.11
What the complements add up to
The 10 complements partition the labour cleanly. Synergetics supplies a metric, a volume convention and an exact packing
in which the carrier’s geometry is computable rather than merely drawable. The FEP supplies a functional and a bound,
and, through predictive coding and the replicator’s reading as inference, reaches two identities – the energy-based learning
gradient and the Shahshahani potential flow – that make the functional computationally and dynamically meaningful. OPH
supplies a finite substrate and an explicit local update that neither of the others has.
Each complement is stated as a numbered proposition: the quadray accounting (Proposition 25), the subdivision identity
(Proposition 26), the learning identity (Proposition 27), the replicator metric (Proposition 28), the carrier’s solid (Proposition
29), the packing (Proposition 30), the three number systems (Proposition 31), the resistance metric (Proposition 32), the
dual pair (Proposition 33) and the conjugation check (Proposition 34).
The synthesis that the complements suggest is therefore not a reduction but an assignment: the carrier is OPH’s, the
functional is the FEP’s, and the accounting is Synergetics’. Two of the complements add a reading neither programme asks
for and both can use: the precision’s own resistance metric, and the conjugation that separates the counts from the volumes.
sec. 9 states the limits of that assignment.
41


## Page 42

Figure 22: Two panels on logarithmic axes. Left: the counting invariants that conjugating the square root of five fixes: the
spanning-tree number 5184000, the automorphism group order 120, the 19 sheaf obstructions and the 20 triangles. Right:
the icosahedron’s tetravolume 18.5123 beside its conjugate 2.700908; the two multiply to the exact norm 50. Generated by
scripts/make_figures.py from geometry/solids.py and the bound values.
9
Scope, limits and falsifiers of a correspondence battery
9.1
What is verified
Everything in sec. 6, sec. 7 and sec. 8 is executed by oph-fep-synergetics verify, which runs 44 probes and exits non-zero
if any of them fails. The command can be narrowed to one pairwise scope with --pair, so a reader who only wants the
Synergetics–OPH ledger does not have to take the rest on trust. The values quoted in the prose are the values the probes
report; a change in any of them changes the probe’s output and, where the change is a failure, changes the exit code.
Four limits of that verification should be stated.
The first is domain. The variational equivalence Proposition 5, the dictionary sec. 6.6 and the two FEP identities hold in the
linear-Gaussian corner. The results are exact there and are not claimed elsewhere. Where a probe can test the boundary of
the domain — the exact-on-domain entries — it does, and the status vocabulary records the difference.
The second is size. The carrier has twelve ports, and the shell census is exhaustive over a set of 78 configurations. Both
are small enough to enumerate and too small to settle any question about scaling. Nothing in this paper claims that the
identities survive a continuum limit, and the obstruction results in particular are statements about this carrier.
The third is provenance. The source for the dynamical controls is a preprint [Mueller, 2026], and OPH’s own results are stated
in its synthesis paper [Mueller et al., 2026]. The preprint’s propositions are cited as the source states them, the package’s
own measurements are reported alongside, and a reader who disagrees with the preprint can still use the measurements. The
three software programmes are pinned to commits, and the pins are point-in-time readings rather than perpetual guarantees.
The fourth is arithmetic. The counting identities are exact: the spanning-tree number is an integer determinant computed
over the rationals, the cohomology dimensions are exact ranks, the reciprocal spectrum sum comes from an exact rational
inverse, the group order and its rotation/reflection split are integers, and the equivariance claim is a finite search over 120
group elements and 78 configurations. The geometric claims are exact too wherever the geometry allows: the tetravolume
table, the cell’s volume and the icosahedron’s tetravolume are computed in the rationals and in ℚ(
√
5) from exact vertex
sets, and the floating-point measurements are reported beside them as a second route. What stays floating-point is what
has to: the eigenvalue locus, the relaxation orbit and the free-energy values of the linear-Gaussian model, whose probes
report residuals and use tolerances stated in the code. The cleanest statement of the two kinds of claim is the conjugation of
√
5, which sec. 8.10 turns into a probe: it fixes the integer counts and moves the volume to 2.700908, and the two volumes
multiply to exactly 50. A counting claim in this paper is therefore invariant under a symmetry that a geometric claim is not,
and comparing one against the other without saying so is the error the section is written to prevent.
42


## Page 43

9.2
Three source quantities and their protocols
The source states three numbers for the native law [Mueller, 2026]. Each is determined by a protocol in OPH’s public model
[FloatingPragma, 2026], and the registry reproduces all three exactly (tbl. 6).
Table 6: The source quantities and the protocols that determine them. sec. 7.1 and sec. 7.6 report the values with their
controls.
Quantity
Protocol
Probe
13/15 and 1/60
one attempt, a uniform seam and a fair coin,
from the configuration with one unit on each of
two adjacent ports: staying put, and one
particular hop
pr01-obstruction
35/264
uniform binary source, two repair events, the
second seam chosen from the first recorded
current: the mean final load at port zero
reduction-criterion
The status not-reproduced stays in the vocabulary for a quantity whose protocol no source states. No entry of the registry
carries it.
9.3
What is not claimed
This paper does not claim that Observer Patch Holography is a free energy principle. It does not claim that the Free Energy
Principle is a holographic theory. It does not claim that Synergetics predicts either. It does not claim that any of the three
is empirically adequate; OPH’s own empirical status document is explicit that no retained laboratory comparison selects a
native record law, and this paper inherits that position rather than improving on it [FloatingPragma, 2026].
It also does not claim novelty for the mathematics. The spectral theory of cellular sheaves, the Routh criterion, the Markov-
blanket critique, and the Goldberg construction are all prior work; what is new here is the conjunction of them on one finite
object, and the discipline of reporting which conjunctions fail.
9.4
Four cautions for anyone continuing this work
The source dossier states four cautions, and this paper adopts them [Friedman, 2026a].
Disambiguate “free energy”. At fixed records, name the minimum interface energy, the unnormalised record potential,
or the normalised record surprisal that bounds variational free energy sec. 7.2.
Respect premise discipline. OPH’s PR-01 describes the native implementation and distinguishes it from the contracts
with consensus theorems. The native and gated relations have different obstructions Proposition 15; a conclusion about one
is not a certificate for the other.
Distinguish numerology from prediction. The twelve-fold and sixty-fold coincidences in this paper are derived, and the
derivation is part of the probe. A numerical coincidence that is not derived is not evidence, and the Synergetics literature
contains both kinds. The counting obstructions are the sharpest examples: twelve-fold features are shared by a cuboctahedron,
a rhombic dodecahedron (twelve faces) and an icosahedron, and only one of the three is the carrier sec. 7.10.
Respect scale separation. The linear-Gaussian results are exact at finite size and say nothing about a continuum limit.
A claim that survives at twelve ports may fail at a thousand, and the paper’s obstruction results are the ones most likely to
change with scale.
9.5
What would change the conclusions
The conclusions are falsifiable in a specific sense, and it is worth saying how.
A single strict descent at |𝑑| = 1 would disprove Proposition 14 and remove the PR-01 obstruction. A pairwise representation
below 𝜀= 0.3333333 would disprove Proposition 19. A member of the solenoidal family with a different entropy production
would disprove Proposition 17. A stable 𝐷𝑘𝐴at 𝑘>
√
10 would disprove Proposition 18. A lattice site whose breadth-
first shell differs from its cuboctahedral gauge would disprove Definition 3, and a transition that fails to commute with an
automorphism would disprove Proposition 10. Each of these is a computation, not an interpretation, and each is executable
against the package as it stands.
43


## Page 44

That is the property this paper was built for: the disagreements are as checkable as the agreements, and either can be
overturned by running a function.
44


## Page 45

10
Reproducibility: pins, commands, and the binding of every number to a
function
10.1
What this manuscript is made of
The manuscript is a project of the docxology/template research monorepo: LaTeX is produced by Pandoc, citations are
resolved with --natbib against references.bib, equations are numbered by pandoc-crossref, and formal statements are
numbered by the repository’s formalism.lua filter, so that a definition or a proposition has the same number in every
rendered edition [Friedman, 2026d].
Every number in the prose is a brace-delimited token resolved at build time from manuscript_variables/values.py, which
binds each value to the function or source that produces it. A token that the generator does not emit fails the build, and a
value that changes changes the prose. The binding is the reason the paper can claim that its numbers are not transcribed.
Table 7: Which surfaces are written by hand and which are produced. A generated surface is never edited to make a check
pass; the producer is changed instead.
Surface
Authored
Generated
manuscript/*.md
yes
never
references.bib
yes
never
src/, geometry/, analysis/
yes
never
../figures/*.png
no
scripts/make_figures.py
../figures/figure_registry.json
no
scripts/make_figures.py
docs/CORRESPONDENCES.md, docs/MANUSCRIPT_VALUES.
md, docs/CONCORDANCE.md
no
scripts/render_docs.py
manuscript tokens
no
scripts/z_generate_manuscript_variables.py
10.2
The three source pins
The three programmes are read at pinned commits. The commit itself lives in data/sources.json and is reported by oph-
fep-synergetics sources, which also reports whether the live checkout still matches it.
Table 8: The three pins. The Recorded column is the number of identifiers the pin file attributes to each source; each
identifier names a file, a line range and a symbol, so that a reader can check the reading rather than trust it.
Source
Repository
Licence
Recorded
Observer Patch
Holography
FloatingPragma/observer-patch-holography
Apache-2.0 (software) and CC
BY-NC-SA 4.0 (papers, data,
ledgers)
commit,
branch,
thirteen cited
identifiers
FEP_Formal
ActiveInferenceInstitute/fep_formal
CC BY 4.0, DOI 10.5281/zeno
do.19699233
commit,
branch, four
cited
identifiers
QuadMath
docxology/QuadMath
Apache-2.0, DOI 10.5281/zeno
do.16887791
commit,
branch, five
cited
identifiers
A pin is a reading, not a promise. oph-fep-synergetics sources reports match when the live checkout is at the pinned
commit, moved when it has advanced, absent when the checkout is not present, and unavailable when Git cannot be
queried. A moved result is not a failure: it says the reading is historical.
10.3
The commands
# the suite, and the coverage gate (90% branch minimum)
.venv/bin/python -m pytest tests/ --cov --cov-report=term-missing
45


## Page 46

# every correspondence, in one run or one pairwise scope at a time
.venv/bin/oph-fep-synergetics verify
.venv/bin/oph-fep-synergetics verify --pair synergetics-oph
.venv/bin/oph-fep-synergetics verify --kind disagreement
.venv/bin/oph-fep-synergetics verify --json
# the pins, checked against the live checkouts
.venv/bin/oph-fep-synergetics sources --vault-root /path/to/vault
# the figures and their accessibility registry
.venv/bin/python scripts/make_figures.py
# the generated reference documents
.venv/bin/python scripts/render_docs.py --check
# the manuscript tokens, then the PDF; both run from the template checkout
cd /path/to/template
TEMPLATE_REPO_ROOT=$PWD .venv/bin/python /path/to/project/scripts/z_generate_manuscript_variables.py
.venv/bin/python scripts/pipeline/stage_03_render.py --project working/OPH-FEP-Synergetics --skip-manuscript-hy
: The reproduction commands, in order. The plain verify runs all 44 probes; --pair runs one ledger, of 26, 14, 2 and 2
claims respectively, and --json writes the same report as machine-readable values. The last two commands run from the d
ocxology/template checkout, because the rendering infrastructure lives there. {#tbl:commands}
10.4
Environment
The package depends on numpy and on nothing else at run time; the test and figure extras add pytest, pytest-cov and
matplotlib. The supported interpreters and library versions are the lower bounds in pyproject.toml (CPython >= 3.12,
numpy >= 1.24), and scripts/build.py prints the interpreter and numpy versions of the run that produced a given PDF.
Every exact value – the counts, the ranks, the tetravolume table, the icosahedron’s volume in ℚ(
√
5) – is computed in Python’s
fractions arithmetic and does not depend on the numpy version. The suite has no network dependency; the provenance
check reads the local checkouts with git rev-parse and reports absent when they are not mounted.
The rendering toolchain is Pandoc 3.13 with pandoc-crossref and a TeX Live installation. The template’s own documen-
tation records the full toolchain [Friedman, 2026d]; this section records only what the manuscript needs.
10.5
What a reader should check first
A reader who wants to test the paper rather than read it should run three commands. oph-fep-synergetics verify
reproduces every number in sec. 6, sec. 7 and sec. 8 and exits non-zero on any failure. oph-fep-synergetics sources shows
whether the three readings are still current. And pytest shows whether the values the prose quotes are still the values the
code returns, which is the property that makes this paper different from a summary of three literatures.
46


## Page 47

11
References
The bibliography is references.bib, read by Pandoc during the PDF render. The build invokes Pandoc with --natbib, so
every [@key] citation in the manuscript is rewritten to the corresponding LaTeX command and resolved against the bib file.
Three entries are provenance records rather than literature: oph, fepformal and quadmath identify the pinned checkouts,
their licences and their DOIs. A reader who wants the exact commits should read data/sources.json or run oph-fep-syn
ergetics sources, which reports the live status of each pin rather than repeating a snapshot here.
The bibliography is validated for syntax and for the required fields of each entry type by the template’s citation validator:
uv run python -m infrastructure.reference.citation.cli validate \
projects/working/OPH-FEP-Synergetics/manuscript/references.bib --strict
The validator path above is template-internal: it resolves inside the docxology/template monorepo and does not exist in
this repository, where the equivalent public check is the repository’s own render contract, scripts/check_render.py. The
pre-digital books are cited the way they can be verified: Synergetics by publisher and year, while the full text of its second
volume is archived, with the address carried in the bibliography.
Miguel Aguilera, Beren Millidge, Alexander Tschantz, and Christopher L. Buckley. How particular is the physics of the free
energy principle? Physics of Life Reviews, 40:24–50, 2022. doi: 10.1016/j.plrev.2021.11.001.
Shun-ichi Amari. Information Geometry and Its Applications. Springer, Tokyo, 2016. doi: 10.1007/978-4-431-55978-8.
Ping Ao. Potential in stochastic differential equations: novel construction. Journal of Physics A: Mathematical and General,
37(3):L25–L30, 2004. doi: 10.1088/0305-4470/37/3/L01.
Kenneth J. Arrow and Maurice McManus. A note on dynamic stability. Econometrica, 26(3):448–454, 1958. doi: 10.2307/
1907624.
Martin Biehl, Felix A. Pollock, and Ryota Kanai. A technical critique of some parts of the free energy principle. Entropy, 23
(3):293, 2021. doi: 10.3390/e23030293.
Andries E. Brouwer, Arjeh M. Cohen, and Arnold Neumaier. Distance-Regular Graphs. Ergebnisse der Mathematik und
ihrer Grenzgebiete (3). Springer, Berlin, 1989. doi: 10.1007/978-3-642-74341-2.
Jelle Bruineberg, Krzysztof Dołęga, Joe Dewhurst, and Manuel Baltieri. The emperor’s new markov blankets. Behavioral
and Brain Sciences, 45:e183, 2022. doi: 10.1017/S0140525X21002351.
D. L. D. Caspar and A. Klug. Physical principles in the construction of regular viruses. Cold Spring Harbor Symposia on
Quantitative Biology, 27:1–24, 1962. doi: 10.1101/SQB.1962.027.001.005.
Andy Clark. Whatever next? predictive brains, situated agents, and the future of cognitive science. Behavioral and Brain
Sciences, 36(3):181–204, 2013. doi: 10.1017/S0140525X12000477.
John H. Conway and Neil J. A. Sloane. Low-dimensional lattices. VII. coordination sequences. Proceedings of the Royal
Society of London. Series A, 453(1966):2369–2389, 1997. doi: 10.1098/rspa.1997.0126.
John H. Conway and Neil J. A. Sloane. Sphere Packings, Lattices and Groups. Springer, New York, 3 edition, 1999. doi:
10.1007/978-1-4757-6568-7.
H. S. M. Coxeter. Regular Polytopes. Dover, New York, 3 edition, 1973.
Lancelot Da Costa, Thomas Parr, Noor Sajid, Sebastijan Veselic, Victorita Neacsu, and Karl Friston. Active inference on
discrete state-spaces: a synthesis. Journal of Mathematical Psychology, 99:102447, 2020. doi: 10.1016/j.jmp.2020.102447.
Lancelot Da Costa, Karl Friston, Conor Heins, and Grigorios A. Pavliotis. Bayesian mechanics for stationary processes.
Proceedings of the Royal Society A, 477(2256):20210518, 2021. doi: 10.1098/rspa.2021.0518.
A. P. Dempster. Covariance selection. Biometrics, 28(1):157–175, 1972. doi: 10.2307/2528966.
Amy C. Edmondson. A Fuller Explanation: The Synergetic Geometry of R. Buckminster Fuller. Birkhäuser, Boston, 1987.
doi: 10.1007/978-1-4684-7485-5.
P. J. Federico. Descartes on Polyhedra: A Study of the De Solidorum Elementis. Sources in the History of Mathematics and
Physical Sciences. Springer, New York, 1982. doi: 10.1007/978-1-4612-5759-2.
Miroslav Fiedler. Algebraic connectivity of graphs. Czechoslovak Mathematical Journal, 23(2):298–305, 1973. doi: 10.21136
/CMJ.1973.101168.
47


## Page 48

Chris Fields, Karl Friston, James F. Glazebrook, and Michael Levin. A free energy principle for generic quantum systems.
Progress in Biophysics and Molecular Biology, 173:36–59, 2022. doi: 10.1016/j.pbiomolbio.2022.05.006.
FloatingPragma.
Observer patch holography.
Git repository, https://github.com/FloatingPragma/observer-patch-
holography, 2026. Software, including the Lean development and the model code, is Apache-2.0; papers, documentation,
figures, data and ledgers are CC BY-NC-SA 4.0.
Daniel Ari Friedman. QuadMath. Zenodo, 2025. Quadray coordinates, the A3/IVM lattice, omni-numbering and exact
tetravolumes. Apache-2.0.
Daniel Ari Friedman. Observer patch holography, the free energy principle, and synergetics: a research dossier. Unpublished
working dossier, 2026a.
Daniel Ari Friedman. FEP_Formal. Zenodo, 2026b. Concept DOI, resolving to the latest version; the commit read is
pinned in data/sources.json. A standalone catalogue of the Free Energy Principle, Active Inference, Bayesian mechanics,
information geometry and thermodynamics, with Lean 4 source. CC BY 4.0.
Daniel Ari Friedman. Observers, blankets, and screens: an integration proposal. Unpublished working proposal, 2026c.
Daniel Ari Friedman. A template/ approach to reproducible generative research: Architecture and ergonomics from configu-
ration through publication. Zenodo, 2026d. The rendering and validation infrastructure this manuscript is built with.
Karl Friston. The free-energy principle: a unified brain theory?
Nature Reviews Neuroscience, 11(2):127–138, 2010. doi:
10.1038/nrn2787.
Karl Friston. Life as we know it. Journal of the Royal Society Interface, 10(86):20130475, 2013. doi: 10.1098/rsif.2013.0475.
Karl Friston. A free energy principle for a particular physics. arXiv:1906.10184, 2019.
Karl Friston and Klaas E. Stephan. Free-energy and the brain. Synthese, 159(3):417–458, 2007. doi: 10.1007/s11229-007-
9237-y.
Karl Friston, Thomas FitzGerald, Francesco Rigoli, Philipp Schwartenbeck, and Giovanni Pezzulo. Active inference: a process
theory. Neural Computation, 29(1):1–49, 2017. doi: 10.1162/neco_a_00912.
R. Buckminster Fuller and E. J. Applewhite. Synergetics: Explorations in the Geometry of Thinking. Macmillan, New York,
1975.
R. Buckminster Fuller and E. J. Applewhite. Synergetics 2: Explorations in the Geometry of Thinking. Macmillan, New
York, 1979. URL https://archive.org/details/synergeticsexplo02full.
Carl Friedrich Gauss. Recension der “untersuchungen über die eigenschaften der positiven ternären quadratischen formen” von
ludwig august seeber. Journal für die reine und angewandte Mathematik, 20:312–320, 1840. doi: 10.1515/crll.1840.20.312.
Reprinted from the Göttingische gelehrte Anzeigen of 1831.
Robert Ghrist. Elementary Applied Topology. CreateSpace, Seattle, WA, 2014.
Claude Godrèche and Jean-Marc Luck. Characterising the nonequilibrium stationary states of Ornstein–Uhlenbeck processes.
Journal of Physics A: Mathematical and Theoretical, 52(3):035002, 2019. doi: 10.1088/1751-8121/aaf190.
Michael Goldberg. A class of multi-symmetric polyhedra. Tohoku Mathematical Journal, 43:104–108, 1937.
Thomas Hales, Mark Adams, Gertrud Bauer, Tat Dat Dang, John Harrison, Le Truong Hoang, Cezary Kaliszyk, Victor
Magron, Sean McLaughlin, Tat Thang Nguyen, Quang Truong Nguyen, Tobias Nipkow, Steven Obua, Joseph Pleso, Jason
Rute, Alexey Solovyev, Thi Hoai An Ta, Nam Trung Tran, Thi Diep Trieu, Josef Urban, Ky Vu, and Roland Zumkeller.
A formal proof of the Kepler conjecture. Forum of Mathematics, Pi, 5:e2, 2017. doi: 10.1017/fmp.2017.1.
Thomas C. Hales. A proof of the Kepler conjecture. Annals of Mathematics, 162(3):1065–1185, 2005. doi: 10.4007/annals.2
005.162.1065.
Jakob Hansen and Robert Ghrist. Toward a spectral theory of cellular sheaves. Journal of Applied and Computational
Topology, 3(4):315–358, 2019. doi: 10.1007/s41468-019-00038-7.
Marc Harper. The replicator equation as an inference dynamic. arXiv:0911.1763, 2009.
Josef Hofbauer and Karl Sigmund. Evolutionary Games and Population Dynamics. Cambridge University Press, Cambridge,
1998. doi: 10.1017/CBO9781139173179.
48


## Page 49

John J. Hopfield. Neural networks and physical systems with emergent collective computational abilities. Proceedings of the
National Academy of Sciences, 79(8):2554–2558, 1982. doi: 10.1073/pnas.79.8.2554.
Adolf Hurwitz.
Über die bedingungen, unter welchen eine gleichung nur wurzeln mit negativen reellen theilen besitzt.
Mathematische Annalen, 46(2):273–284, 1895. doi: 10.1007/BF01446812.
Chii-Ruey Hwang, Shu-Yin Hwang-Ma, and Shuenn-Jyi Sheu.
Accelerating gaussian diffusions.
The Annals of Applied
Probability, 3(3):897–913, 1993. doi: 10.1214/aoap/1177005371.
Hassan K. Khalil. Nonlinear Systems. Prentice Hall, Upper Saddle River, NJ, 3 edition, 2002.
Gustav Kirchhoff. Über die auflösung der gleichungen, auf welche man bei der untersuchung der linearen vertheilung gal-
vanischer ströme geführt wird. Annalen der Physik, 148(12):497–508, 1847. doi: 10.1002/andp.18471481202.
Michael Kirchhoff, Thomas Parr, Ensor Palacios, Karl Friston, and Julian Kiverstein. The markov blankets of life: autonomy,
active inference and the free energy principle.
Journal of the Royal Society Interface, 15(138):20170792, 2018.
doi:
10.1098/rsif.2017.0792.
Douglas J. Klein and Milan Randić.
Resistance distance.
Journal of Mathematical Chemistry, 12(1):81–95, 1993.
doi:
10.1007/BF01164627.
Solomon Kullback and Richard A. Leibler. On information and suﬀiciency. Annals of Mathematical Statistics, 22(1):79–86,
1951. doi: 10.1214/aoms/1177729694.
Tony Lelièvre, Francis Nier, and Grigorios A. Pavliotis. Optimal non-reversible linear drift for the convergence to equilibrium
of a diffusion. Journal of Statistical Physics, 152(2):237–274, 2013. doi: 10.1007/s10955-013-0769-x.
Winfried Lohmiller and Jean-Jacques E. Slotine. On contraction analysis for non-linear systems. Automatica, 34(6):683–696,
1998. doi: 10.1016/S0005-1098(98)00019-3.
Beren Millidge, Alexander Tschantz, and Christopher L. Buckley. Predictive coding approximates backprop along arbitrary
computation graphs. Neural Computation, 34(6):1329–1368, 2022. doi: 10.1162/neco_a_01497.
Bernhard Mueller. Agreement and surprise: Local repair, dynamics and the limits of equilibrium. Preprint, Pragma Research
Inc., 22 October 2026. PhilPapers record MUEAAS-2, 2026. URL https://philpapers.org/rec/MUEAAS-2.
Bernhard Mueller, Mario Poneder, Alexander Osika, Kai Xue, Ben Cassie, Peter Nguyen, Jinwook Kim, David Matscheko,
Daniel Ari Friedman, Jonathan Hill, William T. Glynn, Maarten Antonie Visser, Kale Arnav Anirudha, and Brieuc
de La Fournière. Observers who read their universe into existence: Finite consensus and the reconstruction of standard
physics. Preprint. PhilPapers record MUEFOC, 2026. URL https://philpapers.org/rec/MUEFOC.
Thomas Parr, Lancelot Da Costa, and Karl Friston. Markov blankets, information geometry and stochastic thermodynamics.
Philosophical Transactions of the Royal Society A, 378(2164):20190159, 2020. doi: 10.1098/rsta.2019.0159.
Thomas Parr, Giovanni Pezzulo, and Karl J. Friston. Active Inference: The Free Energy Principle in Mind, Brain, and
Behavior. MIT Press, Cambridge, MA, 2022. doi: 10.7551/mitpress/12441.001.0001.
Judea Pearl. Probabilistic Reasoning in Intelligent Systems: Networks of Plausible Inference. Morgan Kaufmann, San Mateo,
CA, 1988. URL https://www.sciencedirect.com/book/monograph/9780080514895.
Fernando J. Pineda.
Generalization of back-propagation to recurrent neural networks.
Physical Review Letters, 59(19):
2229–2232, 1987. doi: 10.1103/PhysRevLett.59.2229.
Michael Robinson. Topological Signal Processing. Springer, Berlin, 2014. doi: 10.1007/978-3-642-36104-3.
Edward John Routh. A Treatise on the Stability of a Given State of Motion. Macmillan, London, 1877.
Håvard Rue and Leonhard Held. Gaussian Markov Random Fields: Theory and Applications. Chapman and Hall/CRC,
Boca Raton, FL, 2005. doi: 10.1201/9780203492024.
Noor Sajid, Philip J. Ball, Thomas Parr, and Karl J. Friston. Active inference: demystified and compared. Neural Compu-
tation, 33(3):674–712, 2021. doi: 10.1162/neco_a_01357.
Benjamin Scellier and Yoshua Bengio. Equilibrium propagation: bridging the gap between energy-based models and back-
propagation. Frontiers in Computational Neuroscience, 11:24, 2017. doi: 10.3389/fncom.2017.00024.
Kurt Schütte and Bartel Leendert van der Waerden. Das problem der dreizehn kugeln. Mathematische Annalen, 125(1):
325–334, 1953. doi: 10.1007/BF01343127.
49


## Page 50

Udo Seifert. Stochastic thermodynamics, fluctuation theorems and molecular machines. Reports on Progress in Physics, 75
(12):126001, 2012. doi: 10.1088/0034-4885/75/12/126001.
Siavash Shahshahani. A new mathematical framework for the study of linkage and selection. Memoirs of the American
Mathematical Society, 17(211), 1979. doi: 10.1090/memo/0211.
Claude E. Shannon. A mathematical theory of communication. Bell System Technical Journal, 27(3):379–423, 1948. doi:
10.1002/j.1538-7305.1948.tb01338.x. Part I; continued in volume 27, number 4, pages 623–656.
Peter D. Taylor and Leo B. Jonker. Evolutionary stable strategies and game dynamics. Mathematical Biosciences, 40(1–2):
145–156, 1978. doi: 10.1016/0025-5564(78)90077-9.
H. F. Verheyen. The complete set of Jitterbug transformers and the analysis of their motion. Computers & Mathematics
with Applications, 17(1–3):203–250, 1989. doi: 10.1016/0898-1221(89)90160-0.
James C. R. Whittington and Rafal Bogacz. An approximation of the error backpropagation algorithm in a predictive coding
network with local Hebbian synaptic plasticity. Neural Computation, 29(5):1229–1262, 2017. doi: 10.1162/NECO_a_00949.
50
