Three ways to make a stable image, three different ledgers — and a theorem saying when a fit forbids nothing, a proposition counting the plane’s two closable worlds, and the conjecture that generalises them
A hologram is a recording. Phase information is frozen onto a plate, and the plate does nothing at all until a light source is played across it. The order is in the plate's past, and the plate is inert.
A mirage is not a recording. Hot air over a road bends light continuously, and the image on the highway exists only while the temperature gradient does. Nothing is stored. The medium is computing the image now, and the moment the road cools, the image is gone.
A kaleidoscope is neither. The beads inside are tumbling at random and never stop; what makes the image symmetric is the mirrors. The order is not in the beads and not in the motion. It is in the shape of the container, and the container permits only certain patterns however the beads happen to fall.
These are three genuinely different sources of order, and the useful thing about them is not the metaphor. It is that each one leaves a different bill. This chapter says what the three bills are, shows that the corpus can produce two of them and knows where the third lives, and states the negative test that follows — a test this corpus needs pointed at its own work as much as at anyone else's.
| source | order lives in | dies when | the ledger | in this corpus |
|---|---|---|---|---|
| Recording | the medium's past | the read-out source is cut | a capacity — how much can be stored | Book 8's territory |
| Throughflow | the dynamics | the energy input stops | a cost — entropy produced | Ch 21 §21.8 |
| Boundary | the symmetry group | the container changes | a count — how many configurations are permitted | Ch 21, Ch 25 |
They are not rival explanations and a real system can run all three at once. What matters is that they are distinguishable, and distinguishable by measurement rather than by preference.
The capacity. This is where the actual holographic principle lives, and it is worth saying clearly because the popular version has buried it. The holographic principle is not a claim that the universe is a projected image. It is a counting statement: the degrees of freedom in a region scale with the area of its boundary rather than its volume, which is Bekenstein–Hawking entropy $S = A/4G$ written as an information bound. There is no reference beam, no photosensitive plate and no projector anywhere in it. A critique of the projector leaves the counting statement entirely untouched.
The cost. Section 21.8 supplies this one for the corpus's own contact form. Restore the deleted heat term and $\alpha_{\mathrm{G}}(\dot\gamma) = \delta Q - T\,dS$, which Clausius makes non-positive, with equality exactly on reversible paths. The quantity $-\alpha(\dot\gamma)/T$ is the entropy produced. An order that is maintained by throughflow has a running bill and the bill can be written down.
The count. Chapter 21 supplies the first one: a closed hexagonal shell needs exactly twelve pentagons, because the rotations preserving the lattice are $\mathbb{Z}[\omega]^\times = \mu_6$, so the elementary disclination is $2\pi/6$, and $4\pi \div (2\pi/6) = 12$. Chapter 25 supplies the general one: the number of distinct closures at size $T$ is $W(T) = d_1(T) - d_2(T)$.
The three ledgers give a diagnostic, and it is sharper than "is this framework rigorous?" — a question nobody can answer about their own work. It is: which of the three bills does this framework produce, and can that bill come back empty?
The kaleidoscope column is the one most often claimed and least often earned, because "the geometry of the container selects the pattern" is easy to assert. The test of whether a selection rule is real is whether it forbids anything. A rule that permits every observation has not selected; it has described.
Let $\mathcal{L}_c = \{\,c/N : N \in \mathbb{Z}^+\}$. For any target $f$ with $0 < f < c$, the nearest element of $\mathcal{L}_c$ satisfies
$$\frac{\bigl|\,\text{nearest rung} - f\,\bigr|}{f} \;\le\; \frac{f}{2c}\,\bigl(1 + O(f/c)\bigr).$$Proof. Put $N_0 = \lfloor c/f \rfloor$, so that $c/(N_0{+}1) \le f \le c/N_0$. The gap containing $f$ has width $$\frac{c}{N_0} - \frac{c}{N_0+1} \;=\; \frac{c}{N_0(N_0+1)} ,$$ and $f$ lies within half of it from one endpoint or the other, so the absolute error is at most $c/\bigl(2N_0(N_0{+}1)\bigr)$. Since $N_0 \le c/f \le N_0 + 1$, we have $N_0(N_0{+}1) \ge (c/f)(c/f - 1) = (c/f)^2\bigl(1 - f/c\bigr)$, whence the absolute error is at most $\tfrac{1}{2}(f^2/c)\bigl(1-f/c\bigr)^{-1}$. Dividing by $f$ gives the stated bound. $\;\blacksquare$
Corollary. For $f \ll c$ the ladder matches every target to arbitrary relative precision. It therefore excludes nothing, and a fit to it carries no information.
Theorem 26.1 is formalised in LadderBound.lean and machine-verified: gap_eq, half_gap, ladder_abs_error, ladder_rel_error, ladder_rel_error_of_lt and ladder_forbids_nothing each depend on [propext, Classical.choice, Quot.sound] and nothing else — no sorryAx, gated by tools/axiom_gate.py at six theorems. The Lean statement is multiplicative, 2 * n * min (…) ≤ f, which is the relative bound cleared of its division.
Which is the chapter’s own standard met by the chapter: §26.4 says a claim that a structure is selected must produce a ledger, and §26.2’s negative test is now the one thing here that a kernel has read.
ch26-verify.py checks the bound numerically against 200,000 random targets in $(1, 200)$ with $c = 700$ and finds zero violations. Concretely, below 50 Hz such a ladder matches any frequency whatever to better than 4%, and near 10 Hz to better than 0.8%. COMPUTED
So a ladder of that shape is not a prediction. It is a curve with one free parameter and effectively unlimited resolution in the band where the claims are made, and finding that a measured frequency “lies on the ladder” conveys no information at all. The correct response to such a fit is not scepticism about the fitter. It is the observation that the family could not have failed.
Chapter 25's $W(T)$ has no free parameter and returns zero on a large set. Below 40 the forbidden sizes are $T = 2, 5, 6, 8, 10, 11, 14, 15, 17, 18, 20, 22, 23, 24, 26, 29, 30, 32, 33, 34, 35, 38$ — no shell of those sizes exists, and the arithmetic says so before any shell is built.
Counted to $x = 200{,}000$, the closing field forbids about 80.6% of sizes, and by Landau–Ramanujan the permitted fraction tends to zero like $1/\sqrt{\log x}$. A one-parameter ladder forbids 0%, at every $x$. COMPUTED
The two sit at opposite extremes of the only property that matters for a selection rule. This is not a difference of rigour, presentation or good faith. It is a difference in the degrees of freedom of the fitting function, and it is measurable.
§26.2 catches one failure: a count that cannot return zero. There is another, and it was not visible from inside this chapter’s examples because both of them — the closing field and the ladder — are evaluated on every instance of their index. Trefethen’s account of P vs. NP supplies the case that shows the gap.
Worst-case complexity is a ledger in the sense of §26.1, and a strict one: it returns a count, it has no free parameter, and it forbids a great deal. Yet the practical force of the P/NP distinction has drained away, and Trefethen gives two reasons. Exponential-in-the-worst-case algorithms often run fast on the instances anybody actually has — the simplex method remains a workhorse of large-scale optimisation despite being provably exponential, and 3-SAT solvers became industrial tools, to the point where Knuth’s 2016 von Neumann Prize lecture was all about SAT applications and mentioned NP-completeness only in passing. And where that fails, accepting slightly less than optimality removes the exponential even in the worst case: a polynomial algorithm gets within 88% on max cut.
A ledger can be exact, parameter-free, and able to return zero, and still forbid nothing that matters — if it is computed over a measure the world does not sample. Worst-case complexity quantifies over all inputs; instances arriving from practice occupy a thin, structured subset. A bound that binds only off the support of what occurs is not a false bound. It is a true one with no purchase.
So the test in §26.2 needs a second clause. Ask not only can this count return zero but does it return zero anywhere the system actually goes. The first question catches a decoration; the second catches a theorem.
Which is worth stating carefully, because it does not license the reading it invites. Nothing here bears on whether P equals NP, and nothing in this corpus does. Trefethen’s observation is his, published under his name, and it is an argument about the consequences of a distinction rather than about the distinction itself — he is explicit that P vs. NP’s “visibility as a pillar of theoretical computer science has only grown”, with 551 complexity classes catalogued as of August 2026. What this chapter takes from it is a diagnostic it was missing, not a result.
And that distinction is the one §26.4 exists to enforce. A framework that can supply a vocabulary for someone else’s finding has an instrument; it does not thereby have their finding, and it certainly does not have the theorem their finding is about. Recognising the shape of a problem is not the same as having solved it, and the gap between those two is where most of the mathematics lives.
Book 8's Chapter 13 names the fourth face. Topology connects the parts, topography maps the surface, holography inscribes the whole — and holology is the logic by which a totality is coherent as a totality. That chapter is the naming, and it closes with an honest limit: what holology is in itself “remains structurally invisible from inside”.
This is the meeting point, and it is exactly the kaleidoscope column. Holology is not a claim about storage and not a claim about drive. It is a claim about constraint — the container's logic deciding what a whole can be. Book 8 named that face and could not see it from inside.
Chapters 21 and 25 are the first place in this corpus where holology is counted. The unit group $\mu_6$ decides which defects can exist; twelve of them are required; and $W(T) = d_1(T) - d_2(T)$ says how many wholes of each size are permitted — returning zero when none is. That is the logic of the whole as a whole, written as an arithmetic function of the system's own index, with no free parameter and the ability to forbid.
Book 8 said the deepest layer was invisible from inside. It was not invisible; it was uncounted. What made it visible was not a better vantage point but a norm form.
Which raises the obvious objection to the join, and the objection has an answer. One instance of a logic is an instance, not the logic. But the closing field is not one instance — it is one of exactly two, and the two exhaust what the plane permits.
Let $\Lambda$ be a planar lattice whose rotation group about a lattice point is $\mu(K)$, the roots of unity of an imaginary quadratic field $K$ with $\mathcal{O}_K$ its ring of integers. There are exactly two such $K$ with $|\mu(K)| > 2$:
| $K$ | $\mathcal{O}_K$ | $|\mu|$ | norm form | count $W(n)$ | defects $2|\mu|$ | realised by |
|---|---|---|---|---|---|---|
| $\mathbb{Q}(\sqrt{-3})$ | $\mathbb{Z}[\omega]$ | 6 | $a^2+ab+b^2$ | $d_1 - d_2 \ (\mathrm{mod}\ 3)$ | 12 | icosahedron: 12 vertices of degree 5 |
| $\mathbb{Q}(i)$ | $\mathbb{Z}[i]$ | 4 | $a^2+b^2$ | $d_1 - d_3 \ (\mathrm{mod}\ 4)$ | 8 | cube: 8 vertices of degree 3 |
In both rows: the number of distinct closed wholes at index $n$, enantiomers counted separately, equals the stated divisor difference — verified with no mismatch below 2000 in each case; the count returns zero on a set of density one; and closure requires exactly $4\pi \div (2\pi/|\mu|) = 2|\mu|$ elementary defects, confirmed against the two polyhedra, both with $V - E + F = 2$. COMPUTED
Every other imaginary quadratic field has $\mu = \{\pm 1\}$. So the two rows are not two samples from a large space. They are the whole of it, for the plane.
The pieces are classical — Jacobi’s two-square theorem, its Eisenstein analogue, Descartes on total defect — and nothing here proves any of them. What is being claimed is only that they are the same statement twice, and that the statement is the one Book 8 was reaching for: the logic of the whole, counted, from the symmetry group of the substrate.
Let a substrate carry a discrete symmetry group $G$ under which its wholes must close. Then:
(i) closure requires exactly $4\pi \div (2\pi/|G|) = 2|G|$ elementary defects;
(ii) the number of distinct wholes at index $n$ is an ideal-counting function of the associated order, hence a divisor sum;
(iii) that function vanishes on a set of density one.
Falsification, as stated: exhibit a substrate whose realisable sizes are not counted by an ideal-counting function, or whose forbidden set has density zero rather than one, or whose defect count differs from $2|G|$. A single counterexample in three dimensions would be enough.
The counterexample arrived the same day, from the direction the falsification clause named, and it is worth reporting in that order — conjecture, then test, then verdict — rather than quietly restating the conjecture as though it had always been narrower.
The natural three-dimensional analogue is not another quadratic ring but the Hurwitz order $\mathcal{H}$, the maximal order in the rational quaternions. Its unit group has order 24 — the vertices of the 24-cell — against $|\mu(\mathbb{Z}[\omega])| = 6$ and $|\mu(\mathbb{Z}[i])| = 4$. Its norm is the quaternary form $a^2+b^2+c^2+d^2$.
(ii) survives, and gains its best instance. By Jacobi’s four-square theorem the representation count is $r_4(n) = 8\sigma(n)$ for odd $n$ and $24\sigma(m)$ for even $n$ with odd part $m$ — checked against direct enumeration for $n \le 20$ with no discrepancy. The count is still a divisor sum, and this is the first instance in a non-commutative order. COMPUTED
(iii) is refuted. Lagrange (1770): every non-negative integer is a sum of four squares. So the forbidden set is empty — density zero, not density one. Direct check to $n = 500$ finds no $n$ with $r_4(n) = 0$. Against the planar cases, which forbid 74.0% ($\mathbb{Z}[\omega]$) and 69.0% ($\mathbb{Z}[i]$) of the first two thousand sizes, the Hurwitz order forbids 0%. COMPUTED
(i) was under-stated, not wrong. The $4\pi$ is the sphere’s $2\pi\chi$. Restore the Euler characteristic and Descartes gives the general law directly:
$$\text{total defect} = 2\pi\chi(\Sigma), \qquad \text{elementary defect} = \frac{2\pi}{|G|}, \qquad \boxed{\;\#\text{defects} = |G|\cdot\chi(\Sigma)\;}$$which reduces to $2|G|$ exactly when $\chi = 2$. It does not generalise into a third dimension — it generalises across surfaces, which is the direction it was always pointing.
The sphere is a bad witness for a law about $\chi$. On the sphere $\chi = 2$, so $|G|\chi$ and $2|G|$ agree, and a constant that happens to equal two is indistinguishable from a coincidence. Every classical statement of the twelve-pentagon theorem is stated on the sphere, which is why the $\chi$ was never visible in it: there was nothing for it to vary against.
To see a law you need the quantity to move. Three things have to be understood together, and each is easier watched than read, so each has a figure below: what a single defect is and why its size is fixed by $|G|$; what non-orientability actually does to a surface, which is not a picture but a process; and how the two combine into a count.
The wedge is the whole mechanism, and it is why $|G|$ enters at all. The rotation that carries one cut edge onto the other must be a symmetry of the lattice, so the wedge angle is a multiple of $2\pi/|G|$ and nothing finer. A triangular sheet pays in sixtieths of a turn, a square sheet in quarters. The lattice fixes the denomination of the currency; the surface fixes the bill.
And the bill is Descartes’. The total angular deficit of any closed surface built this way is $2\pi\chi(\Sigma)$ — a purely topological quantity that knows nothing about lattices. Divide the one by the other and the defect count falls out with no freedom left in it.
Which brings the second thing, and it is the one that cannot be drawn statically. A surface is non-orientable when a frame carried around some loop comes back reversed. That is a statement about a journey, not about a shape, and the Möbius band is the smallest place it happens.
Nothing in that process costs curvature. The band is flat — it can be made from paper without stretching — and the frame’s reversal is a global fact about how the strip is glued, not a local one about how it bends. That is the first hint of the result in §26.3’s last box: non-orientability is a real structural property that is free in the currency Chapter 21 counts.
Now put the two together. The lattice sets the size of one defect; the surface sets the total; and the count is forced.
The table below is the same statement as the figure, in numbers, and the non-orientable rows are the exemplary ones: they are where the sphere-specific version fails outright and the general version keeps working.
| surface | $\chi$ | orientable | $|G|$ | $|G|\chi$ predicted | realised as |
|---|---|---|---|---|---|
| sphere | 2 | yes | 6 | 12 | icosahedron — 12 vertices of degree 5 |
| sphere | 2 | yes | 4 | 8 | cube — 8 vertices of degree 3 |
| torus | 0 | yes | 6 | 0 | the hexagonal sheet tiles it flat, no pentagons |
| $\mathbb{RP}^2$ | 1 | no | 6 | 6 | hemi-dodecahedron — 6 pentagons |
| $\mathbb{RP}^2$ | 1 | no | 4 | 4 | hemi-cube — 4 vertices of degree 3 |
| Klein bottle | 0 | no | 6 | 0 | closes with no defect at all |
Five predictions, five confirmations, two of them on non-orientable surfaces, and none of them adjustable. COMPUTED The hemi-polyhedra are the cleanest: $\mathbb{RP}^2$ is the sphere quotiented by the antipodal map, $\chi$ halves from 2 to 1, and the defect count halves with it — the icosahedron’s twelve pentagons become the hemi-dodecahedron’s six, the cube’s eight corners become the hemi-cube’s four. A law stated as “$2|G|$” predicts twelve and six respectively, and is simply wrong on both. The law stated as $|G|\chi$ is right on all five.
The Klein bottle has $\chi = 0$. A hexagonal sheet closes onto it with zero defects. So Klein topology is not expensive; it is free, in exactly the currency Chapter 21 counts. Any framework that invokes a Klein or Möbius boundary condition as the thing which selects a structure is, on this ledger, invoking something that costs nothing and therefore forbids nothing — the same objection §26.2 makes to a one-parameter ladder, arriving from topology instead of arithmetic.
Non-orientability may still do real work elsewhere — it splits mode spectra in ways orientable cavities cannot, which is a claim about eigenvalues and not about defects. But it does not pay a defect bill, and a selection argument has to say which bill it is paying. OPEN
What survives is narrower and, being narrower, is worth more: the count is an ideal-counting function — verified now in three orders, two commutative and one not — together with the defect law $|G|\chi(\Sigma)$, verified on five surfaces. Clause (iii) is gone, and nothing more general than these two has been established.
The refutation has a reading, and it inverts the intuition that made the conjecture attractive. The planar cases forbid a great deal because their unit groups are small — four and six — and their norm forms are binary. The Hurwitz order forbids nothing because it has twenty-four units and a quaternary form. A count’s power to forbid comes from the scarcity of its symmetry, not from its richness.
Which runs in the same direction as WP-103: a lattice admits a rotation of order $n$ when $\varphi(n) \le d$, so higher dimensions permit more and therefore forbid less. A framework that reaches for a larger symmetry group in the hope of a stronger selection rule has the sign backwards. This chapter nearly did.
The logo of IMPA — Brazil’s Instituto de Matemática Pura e Aplicada, founded 1952 as the first research unit of the civilian research council CNPq — is a Möbius band. An institute whose emblem is the simplest non-orientable surface sits in the country whose school curriculum this corpus discusses elsewhere, and the surfaces that made the defect law visible above are the non-orientable ones.
That is a pleasing convergence and it is not evidence of anything. WP-29’s rule binds here as everywhere: a shared object between two domains is not a bridge until the mechanism is shared, and there is no mechanism connecting an institute’s visual identity to the Euler characteristic of its country’s tilings. It is recorded because noticing a resonance and then declining to spend it is the habit this chapter is about.
A diagnostic written to evaluate other people's frameworks and never turned inward is not a diagnostic; it is a weapon. WP-29 established the habit for this corpus and it applies here first.
Where this corpus passes. Chapter 21's twelve is a count that forbids — it says eleven and thirteen are impossible. Chapter 25's $W$ forbids about four sizes in five. Section 21.8's entropy production is a cost with a sign fixed by Clausius rather than by us. Three ledgers, two of them ours, both able to come back empty.
Where it does not. The corpus's contact form $\alpha = dz - r^2 d\theta$ was written for a hundred and sixty-four files before anyone noticed a term was missing, and the entropy §21.8 asked for has been named for exactly one system. The $700/N$ objection in §26.2 should be read against our own $\eta$-ladder and $\varphi$ material with the same eye: wherever this corpus fits a constant to a phenomenon, the question is what that fit forbids, and the answer must be checkable.
The rule, stated so it binds us. Any claim in this corpus that a structure is selected must come with one of the three ledgers, and if it comes with a count, that count must be able to return zero. Claims that cannot meet this stay marked OPEN — which is where several of them currently are, and where they should stay until they can.
26.1 Show that a two-parameter family $\{a/N + b\}$ fits any finite set of targets exactly, and conclude that counting free parameters is the first thing to do to any proposed selection rule.
26.2 The bound in Proposition 26.1 degrades as $f \to c$. At what $f$ does a $700/N$ ladder begin to forbid something, and what does that say about where such a claim could in principle be tested honestly?
26.3 Which of the three ledgers does the Caspar–Klug classification of viral capsids produce? Justify the answer by naming what it forbids.
26.4 Conjecture 26.3(iii) says the forbidden set has density one. Verify this for $\mathbb{Z}[i]$ directly — the realisable sizes are the sums of two squares, whose counting function is $\sim K x/\sqrt{\log x}$ (Landau). Which classical theorem is the $\mathbb{Z}[\omega]$ analogue?
26.5 Take any chapter of this corpus that claims a structure is selected and apply §26.4's rule to it. Name the ledger or mark the chapter open.
sorryAx. A clean axiom report is not a reading of the statement: per R20, a theorem can assume its conclusion and still report clean. Follow the link before citing one as evidence.