Layer 1 — Why “simplest” means description-length — rendered package. Rendered from layer-1-metric.md; frozen technical content unchanged by rendering.

Layer 1 — Why “simplest” means description-length

Whether 13 dimensions is “simpler” than 4 depends entirely on how you measure simplicity — and that choice is the decisive seam of the whole program.

This is the first bridge in the stack that runs from the master anchor down to a concrete claim about the 13D shape. It is also the most surprising one, because it shows that the obvious question — “how can a 13-dimensional theory possibly be simpler than a 4-dimensional one?” — is the wrong question. The honest answer is: it depends on the ruler. Pick the ruler that counts dimensions, and a clean four-dimensional effective field theory wins by definition, since $4 < 13$. Pick the ruler that counts bits of brute fact you have to feed in by hand, and the 13D branch wins — on the scored ledger below. Layer 1 is about which ruler the universe actually hands you — and exactly how far we have gotten in proving it.

The question this layer answers

When two theories both reproduce the Standard Model, which one is “simpler”? That word has no meaning until you fix a metric. The entire force of “13D is minimal” lives or dies on that single choice. Layer 1’s job is to identify the right metric, derive it rather than assume it, and say plainly what part of that derivation is finished and what part is still an open theorem.

The headline result is a genuinely positive one, and it deserves to be stated confidently: under the metric that a finite-record universe actually forces — description length — the 13D branch is the shortest complete recipe of every competitor we have scored, and the reason it wins is non-circular. But the metric itself rests on one explicit interpretive bridge, and that bridge is what keeps the absolute claim open. We say exactly where that line is.

Two rulers, two winners

Start with the fork in its sharpest form. Write any candidate theory $B$ as a recipe, and ask for its cost. There are two honest ways to count.

Ruler 1 — dimension-first. Order candidates lexicographically with the spacetime+internal dimension count $k_{\rm dim}(B)$ leading:

$$ C_{\rm dim}(B) = k_{\rm dim}(B). $$

Under this ruler a four-dimensional chiral-gauge effective field theory (the Standard Model written directly as a 4D quantum field theory) beats the 13D branch on the very first key, $4 < 13$, irrespective of anything else — irrespective of how many measured numbers it has to inject by hand. Under dimension-first, minimality folds for everyone who builds in extra dimensions, full stop.

Ruler 2 — description length (MDL). Charge a candidate by the number of bits needed to specify its brute-fact inputs well enough to reproduce the target to the resolution at which we can actually measure it:

$$ C(B) = I(B) = I_{\rm struct}(B) + I_{\rm gen}(B) + n(B)\,b + I_{\rm extra}(B), $$

where $I_{\rm struct}$ is the cost of discrete structural choices (which dimension, which coset, which quotient, which integer index), $I_{\rm gen}$ is the cost of the generator — the machinery that maps inputs to observables — $n(B)$ is the number of independently-measured real numbers the candidate must inject, $b = \log_2(1/\Delta_0)$ is the bit-cost of pinning down one such real to the operational resolution $\Delta_0$, and $I_{\rm extra}$ collects any further charged rules. Under this ruler, injected reals are expensive and structural choices are cheap, so a theory that generates a number can beat a theory that hand-tunes it — even if the generating theory carries more dimensions.

Both rulers are internally consistent. They simply rank the same two theories oppositely. That is the whole problem. The question is not “which theory is simpler” but “which ruler is correct,” and that has to be answered by a principle, not a preference. Choosing the ruler that happens to make your favorite theory win would be the cardinal sin of this kind of work — it is the move we most carefully refuse.

The principle that picks the ruler: finite operational reality

Here is the architecture-neutral argument, and it is the heart of Layer 1.

The Finite Operational Cell Law. There exists a resolution $\Delta_0 > 0$ such that physically distinguishable records occupy finite cells. A measurement does not return an infinite-precision real number; it returns a cell. Consequently a candidate theory is not specified by infinite-precision continuum data — it is specified by a finite operational record string. The universe, as far as anyone can ever interrogate it, is a finite-record system.

Now ask: what is the natural complexity measure for a finite-record system? Impose five conditions on any candidate complexity functional $C(B)$ — each one is a property you would demand of any honest cost accounting, with no reference to dimensions or geometry:

  1. Record faithfulness. $C(B)$ measures the cost of specifying the brute facts needed to reproduce the target to resolution $\Delta_0$.
  2. Encoding invariance. Changing notation changes $C$ by at most an $O(1)$ coding constant.
  3. Additivity. The costs of independent injected anchors add.
  4. Precision monotonicity. Specifying a real number more precisely costs more.
  5. No free hidden data. Fitted tables, tuned normalizations, exception lists, and hidden generators are all charged as information.

Theorem (granularity ⇒ MDL among operational record-cost metrics). Under these five axioms, the unique complexity functional, up to an $O(1)$ coding constant, is description length:

$$ C(B) = I(B) = \min_{p:\,U(p) = T_{\Delta_0}} \ell(p) + O(1), $$

the length of the shortest prefix-free program $p$ that reproduces the target $T$ to operational resolution $\Delta_0$ on a reference machine $U$. The proof is the standard route: a finite-record universe cannot contain or specify infinite-precision continuum data, so its admissible descriptions are finite bit-strings, and “simplest” = “shortest such string” = MDL; the five axioms pin the functional uniquely up to coding.

The crucial corollary — and the reason the fork resolves — is the cost asymmetry between reals and structure:

$$ \text{a measured real anchor costs } \Theta\!\big(\log(1/\Delta_0)\big) = \Theta(b) \text{ bits, large;} $$ $$ \text{a discrete structural choice (dimension, coset } K_6=SU(3)/T^2,\ \text{a } \mathbb{Z}_6,\ \text{an integer index) costs } O(1) \text{ bits, small.} $$

A real number specified to the cell resolution requires you to say which of roughly $1/\Delta_0$ cells it lives in — that is $\log_2(1/\Delta_0)$ bits, and $\Delta_0$ is tiny, so $b$ is large. A choice among a handful of named cosets, or the value of a small topological integer, costs a constant number of bits no matter the resolution. Therefore anchor-burden dominates dimension-burden: many injected reals cost far more than a few extra finite-dimensional structural labels. This is the exact opposite of the dimension-first lex order, which put $k_{\rm dim}$ first by fiat.

Why dimension-first is inadmissible as a record-cost metric

The asymmetry immediately disqualifies the dimension-first ruler — not as a matter of taste, but as a matter of what it claims to measure. Consider two candidates charged against the same target:

$$ B_1 = \text{4D} + 25 \text{ measured reals}, \qquad B_2 = \text{13D} + 13 \text{ measured reals}. $$

Dimension-first declares $B_1$ simpler because $4 < 13$. But operationally $B_1$ needs $\sim 25\,b$ bits and $B_2$ needs $\sim 13\,b + O(1)$ bits, and for $b$ large, $25\,b \gg 13\,b$. So dimension-first ranks as “simpler” the architecture that demands more finite records to pin down. It violates record faithfulness; it is not a record-cost metric at all. It survives only as a non-operational aesthetic prior — a stated preference for fewer dimensions regardless of injected information — which is a legitimate thing to like, but not a measure of how much brute fact a theory contains.

This is what lets the metric scoring of the entire competitor ladder go forward on a principled footing: under the granularity-induced description-length metric, every candidate is charged by the same functional, and dimension-first is off the table as a record-cost ruler.

What this buys, concretely: the economy is real — and strictly counted

With MDL as the operative metric, the 13D branch’s advantage becomes a measured quantity, not a slogan. The branch wins precisely where it converts something a competitor must buy with an injected real into a forced integer or a closed relation:

Every competitor we scored loses by injecting tuned continuous parameters that carry no forced relation — split-fermion localization moduli, a $CP^2$ flux “dialed to 3,” heterotic Yukawas-as-functions-of-stabilized-moduli, membrane-instanton cycle volumes, $\chi^2$-fit $O(1)$ coefficients, finite-Dirac matrix entries, bare lattice Lagrangian Yukawas. Each tunable match is, under MDL, an injected real and therefore a longer recipe. Across the considered ladder of dimensional rungs $D = 4..12$ plus two non-dimensional presentations, the tally is 0 refuted, 1 structural failure-to-generate (the 6D rung, where no 2-manifold isometry contains $SU(3)_c$), and 10 losses to 13D under the MDL metric — no rung yields a strictly shorter complete recipe.

But honesty cuts both ways, and the second discipline line matters here: given-E is not derivation-of-E. The economy is real, and it is strictly counted — two counts answering two different questions, both published (the overview page prints the full two-count ledger). The calibration count answers “which measured values fix the dials?”: from the two flavor anchors $\{y_t, |V_{us}|\}$ the geometry returns 19+ flavor observables, with 6–8 more following from $\{M_{\rm Pl}, \alpha_i\}$ — a handful in, 20+ out. The strict whole-construction count answers “what does the whole machinery run on?”, and it is the count this page charges. The 13D branch’s honest charged cost is

$$ I(B_{13}) \approx (4 + 9\text{–}10)\,b \;=\; 13\,b\text{–}14\,b, $$

i.e. four measured anchors $\{M_{\rm Pl}, \alpha_i, y_t, |V_{us}|\}$ plus roughly nine or ten further injected reals — the fitted normalizations $N_d, N_e, N_\nu$, the $\eta_{BK}$ and $\theta_H^*$ entries, and the threshold triple $\delta$. Compared against a plain 4D effective field theory at $I(B_{\rm EFT}) \approx 25\,b$ (the EFT injects the full flavor ledger directly), the genuine saving is the $\sim 9\text{–}10$-real flavor-hierarchy-plus-unification economy — the 13D recipe runs at close to half the EFT’s bill. Counted strictly across the whole construction, roughly 22 independent full-precision outputs stand against $\sim$5–6 effective inputs — the labeled $\sim$4× over-determination metric — with the complete charged bill above printed beside it. Several 13D outputs are themselves fitted injections; the branch is not treated as free, and that is exactly what the strict count is for. Neither count retires the other: the calibration count is where the predictions live; the strict count is the audit that keeps it honest.

The non-circularity that makes the win count

The reason this is a result and not a tautology is a single bridge that holds the whole ladder together:

A tunable match cannot be counted as a prediction; the tuned value must be injected as a measured real; an injected measured real costs $\sim b$ bits; therefore a tunable match lengthens the description. Anti-fitting $\equiv$ MDL — exactly when the complexity metric is description-length.

The eliminations in the ladder are made by an anti-fitting rule — “if the geometry delivers a constraint only as an adjustable number, that is a fail.” Independently, the MDL codebook charges a fitted normalization at roughly its number of entries times $b$ — i.e. like an anchor. These are the same statement. So the win is adjudicated by the description-length codebook applied identically to the 13D branch and to every competitor, with the observed spectrum $E$ and all of its automatic consequences (anomaly cancellation, written-spectrum chirality, the $\mathbb{Z}_6$ center-kernel, accidental proton stability) cancelled on both sides. Charging a competitor for an $E$-automatic fact would be the same accounting sin run in reverse — as dishonest as under-charging the 13D generator — and we do not do it.

Status — reached vs open

This is where the three discipline lines do their work.

Reached (and stated confidently). - Granularity ⇒ MDL among operational record-cost metrics is proven, conditionally. Under the five record-cost axioms, description length is the unique complexity functional up to $O(1)$, and the real-vs-structure cost asymmetry $\big(\Theta(\log(1/\Delta_0))$ vs $O(1)\big)$ follows. Dimension-first is inadmissible as a record-cost metric — it ranks the more record-hungry theory as simpler. - Under that metric, the 13D branch is the shortest complete recipe across the considered competitor ladder, and the anti-fitting $\equiv$ MDL bridge makes that win non-circular rather than a restatement of the selector’s own preferences.

Open (the named target). The conditional is the seam. Granularity selects MDL among operational record-cost metrics, but the cost-floor alone does not logically forbid every non-MDL preorder. Someone may still define “simplicity” as a separate aesthetic prior — “fewer dimensions first, regardless of injected reals” — and the cost-floor only shows that such a ruler is not an operational record-cost metric. Closing the absolute gap requires one extra, explicit bridge axiom:

$$ \boxed{\ \text{physical simplicity} \;=\; \text{minimal injected operational record cost}\ } $$

This bridge is the natural reading of granularity — the reading a finite-record universe invites. But it is an added interpretive principle, not a consequence of the cost-floor by itself. So the honest verdict on the metric is split: proven (conditional) that granularity uniquely selects MDL among record-cost metrics; not established (absolute) that no aesthetic dimension-first preorder is admissible. Until the bridge axiom is adopted explicitly — or, better, derived — “13D is minimal” remains metric-relative, and a faithful refuted-economy outcome (the generator map turns out not to be cheaper than the reals it replaces, so the simpler 4D realization is the correct one) stays a live and equally valuable possibility. That is not a hedge; it is the falsifier that gives the positive claim its teeth.

And the third discipline line stands over all of it: frozen-and-reproducible is not the same as proven-unique. The 13D branch is hash-frozen and scored blind by the same codebook as every rival — but being the cheapest recipe in the considered class under the operative metric is not a proof that no shorter recipe exists anywhere. Exact MDL over arbitrary programs is uncomputable (the Kolmogorov wall), so absolute minimality is not a theorem anyone can ever close. Selection is not derivation. Layer 1 does not pretend otherwise; it makes the boundary precise, which is exactly what an honest framework owes its reader.

The strongest true sentence for this layer: under the description-length metric that a finite-record universe forces, the 13D branch is the shortest known complete recipe of the data given $E$, the metric itself is uniquely selected by granularity among operational record-cost rulers, and the one remaining move is to promote the simplicity-equals-record-cost bridge from a natural interpretation to a derived principle.

Cross-links